CN117728386A - Prediction method for renewable energy power generation capacity of green building - Google Patents
Prediction method for renewable energy power generation capacity of green building Download PDFInfo
- Publication number
- CN117728386A CN117728386A CN202311522411.2A CN202311522411A CN117728386A CN 117728386 A CN117728386 A CN 117728386A CN 202311522411 A CN202311522411 A CN 202311522411A CN 117728386 A CN117728386 A CN 117728386A
- Authority
- CN
- China
- Prior art keywords
- power generation
- wind power
- prediction
- component
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000010248 power generation Methods 0.000 title claims abstract description 275
- 238000000034 method Methods 0.000 title claims abstract description 126
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 51
- 230000007774 longterm Effects 0.000 claims abstract description 17
- 238000003062 neural network model Methods 0.000 claims abstract description 9
- 230000006870 function Effects 0.000 claims description 56
- 239000013598 vector Substances 0.000 claims description 55
- 230000002068 genetic effect Effects 0.000 claims description 40
- 210000002569 neuron Anatomy 0.000 claims description 38
- 230000008569 process Effects 0.000 claims description 29
- 230000007787 long-term memory Effects 0.000 claims description 28
- 238000001228 spectrum Methods 0.000 claims description 26
- 238000012549 training Methods 0.000 claims description 23
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 claims description 21
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 claims description 21
- 230000015654 memory Effects 0.000 claims description 19
- 230000002159 abnormal effect Effects 0.000 claims description 18
- 230000007246 mechanism Effects 0.000 claims description 14
- 238000012545 processing Methods 0.000 claims description 14
- 230000003044 adaptive effect Effects 0.000 claims description 12
- 238000005457 optimization Methods 0.000 claims description 12
- 230000005855 radiation Effects 0.000 claims description 11
- 239000011159 matrix material Substances 0.000 claims description 9
- 210000004027 cell Anatomy 0.000 claims description 6
- 238000012937 correction Methods 0.000 claims description 6
- 238000001514 detection method Methods 0.000 claims description 6
- 238000009499 grossing Methods 0.000 claims description 6
- 230000006872 improvement Effects 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 6
- 238000005215 recombination Methods 0.000 claims description 6
- 230000009467 reduction Effects 0.000 claims description 6
- 230000001629 suppression Effects 0.000 claims description 6
- 230000006399 behavior Effects 0.000 claims description 5
- 238000002790 cross-validation Methods 0.000 claims description 5
- 238000012216 screening Methods 0.000 claims description 5
- 230000006403 short-term memory Effects 0.000 claims description 5
- 238000006467 substitution reaction Methods 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 230000000694 effects Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 230000006798 recombination Effects 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 230000005764 inhibitory process Effects 0.000 claims description 3
- 230000001537 neural effect Effects 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- 230000033228 biological regulation Effects 0.000 claims 1
- 230000008859 change Effects 0.000 abstract description 3
- 238000013135 deep learning Methods 0.000 description 5
- 238000013461 design Methods 0.000 description 4
- 230000003278 mimic effect Effects 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000005611 electricity Effects 0.000 description 3
- 238000005096 rolling process Methods 0.000 description 3
- 238000010276 construction Methods 0.000 description 2
- 230000007423 decrease Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000035772 mutation Effects 0.000 description 2
- 238000013450 outlier detection Methods 0.000 description 2
- WHXSMMKQMYFTQS-UHFFFAOYSA-N Lithium Chemical compound [Li] WHXSMMKQMYFTQS-UHFFFAOYSA-N 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 229910052799 carbon Inorganic materials 0.000 description 1
- 230000001364 causal effect Effects 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000004146 energy storage Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 229910052744 lithium Inorganic materials 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 229910021421 monocrystalline silicon Inorganic materials 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a prediction method of renewable energy generating capacity of a green building, and belongs to the technical field of renewable energy supply of buildings. The implementation method of the invention comprises the following steps: decomposing the wind power signal into three components of a trend component, a fluctuation component and a random component by a VMD decomposition method, and extracting a complex quantitative relation hidden in a historical wind power signal by combining long-term noise caused by forgetting the random component by using an ILSTM model so as to realize effective prediction of wind power generation; and accurately identifying core meteorological factors influencing the photovoltaic by using a MARS regression method, determining the optimal super-parameter combination deduced by the complex signals by using a GA-LSTM combined neural network model, and deducing the change trend of the optimal super-parameter combination, so that the effective prediction of photovoltaic power generation is realized. The method can avoid the problem that high-precision forecast data such as wind speed, wind direction and the like around the building are difficult to obtain, accurately capture the nonlinear power generation characteristics of the photovoltaic and identify the core meteorological factors affecting the nonlinear power generation characteristics, and improve the forecast precision and forecast efficiency of the renewable energy power generation of the green building.
Description
Technical Field
The invention relates to a prediction method of renewable energy generating capacity of a green building, and belongs to the technical field of renewable energy supply of buildings.
Background
The use of wind and solar energy storage renewable energy systems to energy buildings green has become a focus of industry. The building can carry out renewable energy power generation operation by carrying the photovoltaic power generation plate and the wind driven generator on a roof and the like, so that the self energy consumption requirement is met in a more green low-carbon mode. However, renewable energy power generation is characterized by randomness and intermittence, affected by changes in the natural environment. The photovoltaic power generation is related to various weather variables such as sunlight duration, short-wave radiation, temperature, humidity and the like, the output power difference is large in different time periods, and the accurate capture of the nonlinear power generation characteristics and key meteorological influence factors is a difficulty in realizing effective prediction. In wind power generation, wind speed and wind direction are used as key variables affecting the generated energy of the wind power generation, are greatly affected by regional weather factors, surrounding ground surface obstacles of a fan and other environmental factors, and it is difficult to acquire weather forecast data with high space-time precision from a public platform.
Traditional wind-solar power generation prediction methods are often used for meeting scene requirements of large wind power generation fields or photovoltaic power generation fields, future output power prediction is deduced by utilizing weather forecast data and combining weather element-power generation power causal relations obtained by a power curve model or multiple regression of a generator, but the method is not suitable for single building renewable energy source generating capacity prediction with small occupied area. Data-driven machine learning algorithms such as a recent artificial neural network, a particle swarm algorithm, a long-short-time memory model and the like are also applied to the field, but the problems of complex calculation and excessive fitting caused by more incoherent variables easily occurring in front-end input are solved. In order to ensure the safety and stability of power supply in the building operation stage, a model method for fully considering the wind-solar energy generating capacity characteristics of a green building is needed to finish accurate and efficient mid-short-term wind-solar energy generating capacity prediction, so that the scientific planning of the green energy management of the building is realized.
Disclosure of Invention
The invention aims to provide a prediction method of renewable energy power generation of a green building, which combines signal decomposition, feature recognition and a neural network with a hybrid prediction algorithm to realize middle-short term accurate prediction of wind power and photovoltaic power generation of the green building. Decomposing the wind power signal into three components of a trend component, a fluctuation component and a random component by a VMD decomposition method, and extracting a complex quantitative relation hidden in a historical wind power signal by combining long-term noise caused by forgetting the random component by using an ILSTM model so as to realize effective prediction of wind power generation; and accurately identifying core meteorological factors influencing the photovoltaic by using a MARS regression method, determining the optimal super-parameter combination deduced by the complex signals by using a GA-LSTM combined neural network model, and deducing the change trend of the optimal super-parameter combination, thereby realizing effective prediction of photovoltaic power generation. The method can avoid the problem that high-precision forecast data such as wind speed, wind direction and the like around the building are difficult to obtain, accurately capture the nonlinear power generation characteristics of the photovoltaic and identify the core meteorological factors affecting the nonlinear power generation characteristics, and improve the forecast precision and forecast efficiency of the renewable energy power generation of the green building.
The invention aims at realizing the following technical scheme:
The invention discloses a prediction method of renewable energy power generation of a green building, which is characterized in that high-frequency historical wind-solar power generation data are respectively obtained from a data acquisition system of a wind power inverter and a photovoltaic inverter of the green building and are used as sample data sets; weighting weather station data in a specified radius around the building by using an inverse distance interpolation method to obtain weather condition data capable of accurately reflecting the position of the building; removing abnormal values of the historical wind power generation amount and the historical photovoltaic power generation amount through a quartile detection method, and then carrying out data correction through a cubic spline interpolation method; the corrected wind power generation data are used for decomposing wind power signals into trend components, fluctuation components and random components in a dimension reduction mode by utilizing a VMD decomposition method; the corrected photovoltaic power generation data are used for screening key meteorological factors influencing photovoltaic power generation by using a MARS regression method; for wind power generation prediction, three components after wind power signal decomposition are normalized and used as input of an ILSTM model to establish a wind power generation prediction model, and the ILSTM memorizes short-term influence of historical random components on future wind power generation capacity and forgets long-term influence; taking the normalized key meteorological factor historical value and photovoltaic power generation historical value as GA-LSTM input training neural network model, and confirming the optimal super-parameter combination predicted by LSTM photovoltaic signals through GA algorithm; wind power VMD decomposition signals of 1 day and 1 week before prediction are input into a trained ILSTM model, and power generation prediction results of a short term and a middle term of wind power generation are respectively output; the method comprises the steps of respectively outputting the short-term and medium-term power generation quantity prediction results of photovoltaic power generation by inputting the prediction values of key meteorological variables for predicting the current day and the current week in a trained GA-LSTM model; and adding the predicted wind power generation amount and photovoltaic power generation amount data to obtain a high-precision and high-efficiency green building renewable energy power generation amount prediction result, namely, the green building renewable energy power generation amount prediction is realized.
The invention discloses a renewable energy prediction method for a green building, which comprises the following steps:
step one: and collecting the historical data of the wind-solar energy generating capacity of the green building and relevant meteorological factors influencing the photovoltaic generating capacity. And respectively acquiring high-frequency wind power generation capacity historical data and photovoltaic power generation capacity historical data from a data acquisition system of the green building wind power inverter and the photovoltaic inverter. From the photovoltaic power generation mechanism, potential meteorological elements influencing the photovoltaic power generation amount are determined, and through checking with the type of a preset authoritative meteorological product, the selected meteorological index can be ensured to acquire historical observation data of an hour level, and weather forecast data of a county meteorological station where a building is located can be provided by a weather forecast system to serve as an input variable for the prediction of the subsequent photovoltaic power generation amount. On the selection of meteorological stations, a matching area with the building position as the center and the preset distance as the radius is set, and the numerical values of all the meteorological stations in the area are subjected to inverse distance weighted interpolation, so that the weather conditions of the building position are reflected more accurately.
Step 1.1: and collecting historical data of building wind-solar energy generation capacity and relevant meteorological factors influencing photovoltaic energy generation capacity. And respectively acquiring high-frequency wind power generation capacity historical data and photovoltaic power generation capacity historical data from a data acquisition system of the green building wind power inverter and the photovoltaic inverter, wherein the sample size is required to meet the preset requirement of the follow-up neural network model prediction.
Step 1.2: from a photovoltaic power generation mechanism, potential meteorological elements which influence photovoltaic power generation amount are determined to comprise variables such as sunlight duration, short wave radiation, air temperature, dew point temperature, relative humidity, air pressure and the like, the meteorological elements are almost uniformly distributed in county areas, and observation and forecast data of meteorological sites of the county where the building is located are obtained and serve as input variables for follow-up photovoltaic power generation amount prediction.
Step 1.3: on the selection of meteorological stations, a matching area with the building position as the center and the preset distance as the radius is set, and the numerical values of all the meteorological stations in the area are subjected to inverse distance weighted interpolation, so that the weather conditions of the building position are reflected more accurately. The calculation formula of the specific inverse distance weighted interpolation method is as follows:
wherein AD (x, y) is the atmospheric figure value of the green building site obtained by final interpolation, AD is the meteorological observation value of the site in the ith matching zone, d i For the distance between the i-th known point and the interpolation point, p is the interpolation parameter.
Step two: and (3) after finishing the data preprocessing according to the step one, performing feature selection on the wind power generation amount data, and performing factor recognition on the photovoltaic power generation amount data. And carrying out outlier detection and correction processing on the data before prediction on the historical wind power generation amount and the historical photovoltaic power generation amount, detecting outliers through a quartile method, and carrying out n times of spline interpolation processing on the outlier power generation amount data to obtain outlier substitution points. For wind power generation, a nonlinear unstable wind power signal is decomposed into a trend component, a fluctuation component and a random component by utilizing a VMD decomposition method, and the characteristics of wind power generation are clearly displayed in a layered manner. The trend component is equivalent to the representation of long-term fluctuation of wind power generation, the fluctuation component is equivalent to the representation of short-term fluctuation, and the random component is close to a random signal. During the decomposition of the constructed VMD signal, the individual modes are transformed by Hilbert transform State component u k And (t) converting the spectrum into an analytic signal and then into a single-side spectrum, multiplying the single-side spectrum by a central frequency index term to convert the modal spectrum into a corresponding base frequency band, and estimating the bandwidth of the demodulation signal through Gaussian smoothing. For photovoltaic power generation, aiming at the relevant meteorological elements which are preliminarily screened in the first step, a key meteorological factor which has a larger influence on the solar energy resources of the place and the photovoltaic power generation capacity of the building is identified by utilizing a multi-element adaptive regression spline method MARS. The MARS divides the data space into a plurality of subareas, adopts a 'break-by-break' strategy, each area has own basis function, and then gradually accumulates a function formula approaching to nonlinearity, thereby improving the accuracy and efficiency of data fitting. During the MARS operation, in the first stage, the MARS constructs a number of basis functions based on all input interpretation variables; in the second stage, those basis functions that do not contribute to the prediction, i.e. those useless explanatory variables, are deleted by generalized cross-validation GCV. The historical data of the photovoltaic power generation capacity is used as an interpreted variable, all factors which possibly influence the photovoltaic are used as interpreted variables and input into the MARS, and the factors which can effectively predict the power generation capacity are screened to be used as the input of a follow-up GA-LSTM model through a GCV mechanism of the MARS.
Step 2.1: and carrying out abnormal value detection and correction processing on the data before prediction on the historical wind power generation amount and the historical photovoltaic power generation amount. The detection process by the quartile method comprises the following steps: all data are arranged according to the order of magnitude, and the quartile value is marked as Q 1 I.e. only 1/4 of all data is greater than Q 1 The lower quartile value is Q 2 I.e. only 1/4 of all data is smaller than Q 2 The upper bound is (Q) 1 +1.5(Q 1 -Q 2 ) Lower bound is (Q) 2 -1.5(Q 1 -Q 2 ) Normal observations between the upper and lower bounds, otherwise abnormal values; after the abnormal value is removed, carrying out n times of spline interpolation on abnormal generating capacity data, and setting f (t) as a time interval [ t ] 1 ,t 2 ]The continuous micro-function can remove k abnormal points on the same day, and the generated energy data of 24-k hours are also obtained; let the function f (t) satisfy the condition f (t) =at n +bt n-1 + … +ct+d, in interval [ t ] 1 ,t 2 ]Substituting a normal value in the historical wind power generation data into a second-order continuous derivative, and solving an f (t) expression; the outlier time t is based on the f (t) expression k And f (t) is carried out, and an outlier substitution value is obtained.
Step 2.2: for wind power generation, a nonlinear unstable wind power signal is decomposed into a trend component, a fluctuation component and a random component by utilizing a VMD decomposition method, so that the characteristics of wind power generation are clearly displayed in a layered manner. The trend component may be regarded as a representation of long-term fluctuations of the wind power generation, the fluctuation component may be regarded as a representation of short-term fluctuations, and the random component is close to the random signal. The goal of the VMD is to decompose the real-time input signal f into k subcomponents u with specific sparsity characteristics k . Assuming that the components are compactly centered around their center frequency omega k The bandwidth in the spectrum determines the a priori sparsity of each component. During the process of decomposing the constructed VMD signal, each modal component u is transformed by Hilbert k And (t) converting the spectrum into an analytic signal and then into a single-side spectrum, multiplying the single-side spectrum by a central frequency index term to convert the modal spectrum into a corresponding base frequency band, and estimating the bandwidth of the demodulation signal through Gaussian smoothing. The decomposition process aims at minimizing the sum of the estimated bandwidths of all the components, the solved constraint condition is that the sum of all the components is equal to the original input signal, and the specific variational constraint problem expression is as follows:
s.t.∑ k u k ＝f (3)
wherein u is k (t) is the kth component after decomposition by a variation mode (here k is specifically 3); omega k Center frequency for each component; delta (t) is a dirac function, specificallyf is the original input signal.
To ensure that the accuracy of modal decomposition meets the requirements and that the constraint conditions are strict. Introducing a penalty factor and a Lagrangian multiplier converts the above into an unconstrained optimization problem:
where v is a penalty factor and λ is a lagrangian multiplier. Updating u by iteration k n+1 、ω k n+1 And lambda (lambda) n+1 And obtaining a function optimal solution. Wherein u is k n+1 The corresponding frequency domain is obtained through Fourier transformation, and the specific function expression is as follows:
step 2.3: for photovoltaic power generation, identifying key meteorological factors which have great influence on solar energy resources of the place and photovoltaic power generation capacity of the building by utilizing a multi-element adaptive regression spline method MARS aiming at the relevant meteorological factors preliminarily screened in the step one, wherein the relevant meteorological factors comprise sunlight duration, short wave radiation, air temperature, dew point temperature, relative humidity and air pressure. The relevant meteorological elements are used as the input of the MARS, the building photovoltaic power generation capacity is used as the output of the MARS, and the variables affecting photovoltaic power generation are screened and identified. The MARS divides the data space into a plurality of subareas, adopts a 'break-by-break' strategy, each area has own basis function, and then gradually accumulates a function formula approaching to nonlinearity, thereby improving the accuracy and efficiency of data fitting. The specific equation for constructing the MARS recognition model is as follows:
wherein,represents a basis function, x is an explanatory variable, namely a meteorological element affecting photovoltaic power generation, alpha 0 Is the intercept, a m Is the coefficient of the basis function, M is the basis functionNumber of numbers, K m Is the number of nodes, t km Is the value of a node, s km ∈{-1,+1}。
Each basis function exists in two possible forms:
or alternatively
Wherein a is a constant term, namely a demarcation node of different areas.
The main operation of the MARS process is mainly divided into two stages. In the first stage, the MARS constructs a number of basis functions based on all of the input interpretation variables. In the second stage, those basis functions that do not contribute to the prediction, i.e. those useless explanatory variables, are deleted by generalized cross-validation GCV. The definition of GCV is as follows:
where ζ=m+ 1+p (M/2), p is the penalty factor, p e [2,4].
Therefore, the historical data of the photovoltaic power generation capacity is taken as an interpreted variable, all factors which possibly influence the photovoltaic are taken as interpreted variables and input into the MARS, and the factors which can effectively predict the power generation capacity are screened to be taken as the input of a subsequent GA-LSTM model through a GCV mechanism of the MARS.
Step three: and constructing a prediction model of the renewable energy power generation amount of the green building. The prediction of wind power generation and photovoltaic power generation is based on the basic framework of the LSTM model for unfolding modeling. LSTM is used as deep learning network for learning time series signals, and valuable information is stored in memory neurons through input gates in the learning process, and long-term memory is reserved, so that the LSTM has better recognition and prediction capability on samples with longer distances. Aiming at an ILSTM prediction model of wind power generation design, the improvement is mainly divided into three parts, namely an input vector, a storage unit and an output unit. Taking a trend component, a fluctuation component and a random component which are obtained by the VMD decomposition method in the second step as model input and taking wind power generation amount as output; the random component in the input vector needs to be included as part of the wind power in a short time, but in the long term, due to the unpredictability of the truly random behavior, when the random amount included in the wind power signal is large, the prediction error increases accordingly. Thus, in the improved ILSTM model, the long-term memory of the random component is suppressed, while the short-term memory of the random component is maintained. And (3) based on the principle of a sliding window, forming a sample comprising a plurality of time dimensions by using a trend component, a fluctuation component and a random component obtained by the VMD decomposition method in the step two as an input vector, wherein the output of the prediction model is the predicted value of the wind power generation amount. The prediction model identifies and learns the modes represented in the input vector, and establishes a mapping relation between the input vector and the output vector, so that the prediction capability of the network model is realized. In order to avoid the influence of random components on the long-term memory mode, a parameter vector D= (1, alpha) is added to the input of the component memory unit, wherein the randomness of the wind power signal is evaluated based on the prediction error, and the value of the random component long-term memory unit inhibition parameter alpha is determined. And (3) designing a GA-LSTM prediction model aiming at the photovoltaic power generation, wherein key meteorological factors which influence photovoltaic power generation capacity and are identified by the MARS in the second step are taken as input, and photovoltaic power generation capacity is taken as output. LSTM has numerous superparameters to evaluate in training and predicting data. The super parameters of the model are optimized by introducing a genetic algorithm into the basic LSTM model, so that the average absolute percentage error MAPE in the prediction process is minimized, and a numerical solution is obtained for the problem based on the genetic algorithm GA-AM of the adaptive variation. Genetic algorithms are algorithms that solve unconstrained and constrained nonlinear optimization problems based on natural selection processes that mimic biological evolution. The algorithm iteratively modifies the population of individual solutions. Constructing a plurality of super-parameter combinations from a given super-parameter space based on a genetic algorithm; then each super-parameter combination is brought into a GA-LSTM model, training and predicting are carried out on the data, and an objective function value MAPE is calculated; selecting a super-parameter combination with smaller MAPE (MAPE) as a parent based on the accounting of the first two steps, and continuously performing cross recombination based on the selected super-parameter combination to generate a child; finally, in a given parameter space, the super-parameter combination is randomly adjusted until the best super-parameter combination is found, and super-parameter tuning of the basic LSTM model is completed.
Step 3.1: the prediction of wind power generation and photovoltaic power generation is based on the basic framework of the LSTM model for unfolding modeling. LSTM is used as deep learning network, is specially used for learning time series signals, and is added with memory neuron mechanism based on RNN model, so that it is more suitable for processing long-term information. The LSTM model introduces forgetting gate, input gate and output gate modulation and updates the state of memory neurons.
The first step in the LSTM model is to let the inputs at time t and the outputs of the neurons at time t-1 enter the forget gate, which decides which messages should be removed.
x t Is the input of the moment t, h t-1 Is the output of the neuron at time t-1,and->Is a weight matrix, < >>Is an error vector, sigmoid (z) = (1+e) -z ) -1 Is an activation function.
The second step of the LSTM model is to enter the input at time t and the output of the neuron at time t-1 into the input gate while calculating the candidate value of the neuron state.
Wherein,is a candidate for neuronal status, i t Is an input door, is a->And->Is a weight matrix, < >>And->Is an error vector, +.>
The third step of the LSTM model is to update the state of the neuron by the state of the neuron at the time t-1, the information of the amnestic gate and the input gate screening and the candidate value of the neuron.
Wherein, the ". Is the multiplication of the corresponding elements.
The fourth step of the LSTM model is to output the information output by the gate through the updated neuron state, and generate the output of the neuron at the time t.
h t ＝o t ⊙tanh(c t ) (15)
Wherein,and->Is a weight matrix, < >>Is an error vector o t Is an output door h t Is the output of the neuron state.
Step 3.2: aiming at an ILSTM prediction model of wind power generation design, the improvement is mainly divided into three parts, namely an input vector, a storage unit and an output unit. The trend component, the fluctuation component and the random component obtained by the VMD decomposition method in the second step are required to be used as model input, and the wind power generation amount is required to be used as output; the random component in the input vector needs to be included as part of the wind power in a short time, but in the long term, due to the unpredictability of the truly random behavior, when the random amount included in the wind power signal is large, the prediction error increases accordingly. Thus, in the improved ILSTM model, the long-term memory of the random component is suppressed, while the short-term memory of the random component is maintained.
Based on the principle of a sliding window, the trend component, the fluctuation component and the random component obtained by the VMD decomposition method in the second step are formed into a sample comprising a plurality of time dimensions as input vectors, and the specific form is as follows:
x kt ＝(u k (t-n+1),u k (t-n+2),…,u k (t)) (16)
Wherein x is kt The input vector at the time of the kth component t after VMD decomposition is given, and n is the sliding time window width.
And the output of the prediction model is the predicted value of the wind power generation amount. The prediction model identifies and learns the modes represented in the input vector, and establishes a mapping relation between the input vector and the output vector, so that the prediction capability of the network model is realized. To avoid the effect of random components on the long-term memory pattern, the parameter D is added to the input of the component memory cell:
wherein, D is the suppression vector parameter of the whole three components, D= (1, alpha), alpha is the specific random component long-term memory unit suppression parameter, and the alpha is more than or equal to 0 and less than or equal to 1. The specific meaning of this part of parameters is as follows:
d is designed to be (1, a) so that the first two-tier component can be transferred directly into long-term memory, and the third tier component is multiplied by a before entering the memory cell. Because 0.ltoreq.α.ltoreq.1, the value of the random component is suppressed. The value of alpha depends on the content of random components in the wind power signal. When the randomness is strong, the value of α should be small. For example, if a is designed to be 0, the third layer component with random components will be completely eliminated and not included in long-term memory. When the wind power signal randomness is weak, the value of alpha can be designed to be 1. If α is 1, the third layer component will go entirely to long term memory. Evaluating the randomness of the wind power signal based on the prediction error to determine the value of alpha:
α＝1-err*10 (18)
Wherein,
in the ILSTM, the output gate and output are defined as follows:
o′ t ＝W ox x t +b o (19)
h′ t ＝o′ t ⊙φ(c′ t ) (20)
step 3.3: and (3) designing a GA-LSTM prediction model aiming at the photovoltaic power generation, wherein key meteorological factors which influence photovoltaic power generation capacity and are identified by the MARS in the second step are taken as input, and photovoltaic power generation capacity is taken as output. LSTM has numerous superparameters to evaluate in training and predicting data. The super parameters of the model are optimized by introducing a genetic algorithm into the basic LSTM model, and the following objective function is set for minimizing the average absolute percentage error MAPE in the prediction process:
wherein y is t Is the actual value of the current,is the predicted value, T is the time length, H is the number of hidden units, E is the training number, B is the data batch size, R is the learning rate, F is the learning rate reduction factor, P is the learning rate reduction period, G is the gradient threshold, and Ω is the hyper-parameter space.
Since it is difficult to solve the constrained optimization problem by means of calculus, a genetic algorithm GA-AM based on adaptive mutation is used to solve the problem by means of calculus. Genetic algorithms are algorithms that solve unconstrained and constrained nonlinear optimization problems based on natural selection processes that mimic biological evolution. The algorithm iteratively modifies the population of individual solutions. At each step, the genetic algorithm randomly selects individuals from the current population and uses them as parents to generate next generation offspring. After one generation and another, the population "evolves" into an optimal solution.
The first step of the genetic algorithm is to construct a plurality of superparameter combinations from within a given superparameter space.
The second step of the genetic algorithm is to bring each super-parameter combination into the GA-LSTM model, train and predict the data, and calculate the objective function value, i.e. MAPE.
The third step of the genetic algorithm is to select two sets of better hyper-parameter combinations, i.e. combinations that make MAPE smaller, as parents.
The fourth step of the genetic algorithm is to perform cross-recombination based on the selected hyper-parametric combinations to produce offspring.
And a fifth step of genetic algorithm, wherein the super-parameter combination is randomly adjusted in a given parameter space.
The genetic algorithm continuously repeats the process until the best super-parameter combination is found, and the super-parameter tuning of the basic LSTM model is completed.
Step four: and training a green building renewable energy generating capacity prediction model.
Step 4.1: normalizing historical data of wind power trend components, fluctuation components and random components which are obtained by a VMD decomposition method in the second step and weather factors which influence photovoltaic power generation and are identified by a MARS regression method; dividing the normalized data into a training set, a verification set and a test set;
step 4.2: training and predicting data by adopting a rolling window, and inputting three wind component signals and historical wind power generation data of corresponding time periods into the ILSTM model built in the third step to train a wind power generation prediction model; and (3) inputting relevant historical meteorological data and historical photovoltaic power generation data of a corresponding period into the GA-LSTM model built in the third step to train a photovoltaic power generation prediction model, so as to obtain a trained green building renewable energy power generation prediction model.
Step five: and outputting a prediction result of the renewable energy power generation amount of the green building according to the trained renewable energy power generation amount prediction model of the green building.
Step 5.1: acquiring historical wind power generation amount data of 1 day and 1 week before prediction, decomposing a wind power signal into a trend component, a fluctuation component and a random component according to the second step, acquiring a forecast value of key meteorological factors which influence photovoltaic power generation amount on the current day and the current week, and carrying out normalization processing on related data of wind power generation and photovoltaic power generation amount;
step 5.2: inputting the processed VMD decomposition components into a trained ILSTM model, and respectively outputting wind power generation capacity prediction results of 1 day in the short term and 1 week in the middle term; inputting the processed critical weather factor forecast values into a trained GA-LSTM model, and respectively outputting photovoltaic power generation quantity forecast results of 1 day in the short term and 1 week in the medium term;
step 5.3: and adding the predicted wind power generation amount and photovoltaic power generation amount data to obtain a predicted result of the short-term power generation amount in the renewable energy sources of the green building.
The beneficial effects are that:
1. the invention discloses a prediction method of renewable energy power generation of a green building, which is applied to the technical field of renewable energy power supply of the building and comprises the steps of collecting a historical data set of wind-solar power generation capacity and related meteorological elements, and correcting an abnormal value of the historical power generation capacity; aiming at wind history signals, decomposing the wind signals into trend components, fluctuation components and random components by utilizing a VMD decomposition method, and fully excavating wind power sequence information through dimension reduction; aiming at the photovoltaic historical signals, accurately capturing the nonlinear power generation characteristics of the photovoltaic by using a MARS regression method and identifying key meteorological factors; aiming at the construction of a wind power generation prediction model, three components after wind power signal decomposition are normalized and used as input of an ILSTM model to establish the wind power generation prediction model, and according to the characteristic that a highly-uncertain wind power random component does not have a memory attribute for a long time, the ILSTM model is utilized to forget the random component to memorize for a long time, so that the complex quantitative relation hidden in the historical wind power signal is accurately extracted and learned; aiming at the construction of a photovoltaic power generation prediction model, taking a key meteorological factor historical value and a photovoltaic power generation historical value which are subjected to normalization processing as an input training neural network model of GA-LSTM, adopting a GA-LSTM combined neural network model, determining an optimal super-parameter combination deduced by a complex signal and deducing the change trend of the optimal super-parameter combination deduced by the complex signal; wind power VMD decomposition signals for predicting the previous 1 day and the previous 1 week are input into the trained ILSTM model, and wind power generation capacity prediction results for the short-term future 1 day and the middle-term future 1 week are respectively output; the prediction values of key meteorological variables for predicting the current day and the current week are input into the trained GA-LSTM model, and photovoltaic power generation capacity prediction results of short-term future 1 day and medium-term future 1 week are respectively output; and adding the predicted wind power generation amount and photovoltaic power generation amount data to obtain a short-term power generation amount prediction result in the renewable energy sources of the green building. In summary, the method combines the signal decomposition, the feature recognition and the neural network with the hybrid prediction algorithm and with the power generation feature of the renewable energy sources of the building, and ensures good prediction accuracy under the condition of high medium-short term prediction efficiency.
2. According to the prediction method of the renewable energy power generation capacity of the green building, disclosed by the invention, the wind power signal is reduced in dimension into 3 components by using a VMD decomposition method in combination with the characteristics of the wind power signal, so that the modal mixing problem existing in the traditional decomposition algorithm is avoided, and the decomposition components are ensured to have corresponding physical significance. The back end is connected with the ILSTM model to fully mine long-term and short-term information in the historical time series data, so that the power fluctuation rule of the unit is searched, and the difficulty that high-precision forecast data such as wind speed, wind direction and the like around a building are difficult to obtain is solved. The setting of suppressing the random component long-term memory by the ILSTM model can reduce the prediction error that increases in the case where the prediction duration increases.
3. According to the prediction method of the renewable energy power generation capacity of the green building, disclosed by the invention, the characteristics of the photovoltaic power generation signal are combined, and the core meteorological factors influencing the photovoltaic power generation capacity are identified by using a MARS regression method. Compared with traditional methods such as correlation coefficient, multiple regression and the like, the MARS regression method is more flexible in capturing the nonlinear power generation characteristics of the photovoltaic, so that key meteorological factors can be accurately identified. And factor screening is completed through a MARS regression method, and input data of a rear neural network model are reduced, so that training time of a green building renewable energy power generation quantity prediction model is shortened, and model generalization capability is improved.
Drawings
FIG. 1 is a schematic flow chart of a method for predicting the energy generation capacity of renewable energy sources of a green building.
Fig. 2 is a schematic view of the structure of the ILSTM for a method for predicting the amount of renewable energy generation in green buildings according to the present invention.
FIG. 3 is a schematic diagram of a GA-LSTM method for predicting the amount of energy generated by renewable energy sources in green buildings according to the present invention.
Fig. 4 is a GA schematic diagram of a method for predicting the amount of renewable energy generation in green buildings according to the present invention.
FIG. 5 is a schematic diagram of a rolling window of a method for predicting the amount of energy generated by renewable energy sources of a green building according to the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the following description of the invention will be made with reference to the accompanying drawings and examples.
As shown in fig. 1, the renewable energy prediction method for a green building disclosed in this embodiment specifically includes the following implementation steps:
step one: and collecting historical data of building wind-solar energy generation capacity and relevant meteorological factors influencing photovoltaic energy generation capacity. And respectively acquiring high-frequency wind power generation capacity historical data and photovoltaic power generation capacity historical data from a data acquisition system of the green building wind power inverter and the photovoltaic inverter. From the photovoltaic power generation mechanism, potential meteorological elements influencing the photovoltaic power generation are determined, and through checking with the types of authoritative meteorological products of the national meteorological center, the selected meteorological indexes can be ensured to acquire historical observation data of an hour level, and weather forecast data of a county meteorological station where a building is located can be provided by a weather forecast system to serve as input variables for the prediction of the subsequent photovoltaic power generation. On the selection of meteorological stations, a matching area with a building position as a center and a certain distance as a radius is arranged, and the numerical values of all the meteorological stations in the area are subjected to inverse distance weighted interpolation, so that the weather conditions of the building position are reflected more accurately.
Step 1.1: and collecting historical data of building wind-solar energy generation capacity and relevant meteorological factors influencing photovoltaic energy generation capacity. The high-frequency wind power generation capacity historical data and the photovoltaic power generation capacity historical data are respectively obtained from a data acquisition system of the green building wind power inverter and the photovoltaic inverter, and the sample size meets the preset requirement of the follow-up neural network model prediction and comprises more than 1000 data sets.
Step 1.2: from the photovoltaic power generation mechanism, potential meteorological factors influencing photovoltaic power generation are determined to comprise variables such as sunlight duration, short wave radiation, air temperature, dew point temperature, relative humidity, air pressure and the like, the factors are almost uniformly distributed in county regions, and observation and forecast data of weather sites of the county where the building is located can be obtained on a public platform and used as input variables for the prediction of the follow-up photovoltaic power generation.
Step 1.3: on the selection of meteorological stations, a matching area with a building position as a center and a certain distance as a radius is arranged, and the numerical values of all the meteorological stations in the area are subjected to inverse distance weighted interpolation, so that the weather conditions of the building position are reflected more accurately. The calculation formula of the specific inverse distance weighted interpolation method is as follows:
wherein AD (x, y) is the atmospheric figure value of the green building site obtained by final interpolation, AD i For the weather observation value of the site in the ith matching zone, d i For the distance between the i-th known point and the interpolation point, here set to 30km, p is the interpolation parameter, here set to 2.
Step two: and finishing data preprocessing, performing feature selection on wind power generation data, and performing factor identification on photovoltaic power generation data. And carrying out outlier detection and correction processing on the data before prediction on the historical wind power generation amount and the historical photovoltaic power generation amount, detecting outliers through a quartile method, and carrying out n times of spline interpolation processing on the outlier power generation amount data to obtain outlier substitution points. For wind power generation, a nonlinear unstable wind power signal is decomposed into a trend component, a fluctuation component and a random component by utilizing a VMD decomposition method, so that the characteristics of wind power generation are clearly displayed in a layered manner. The trend component may be regarded as a representation of long-term fluctuations of the wind power generation, the fluctuation component may be regarded as a representation of short-term fluctuations, and the random component is close to the random signal. In the process of decomposing the constructed VMD signal, each modal component u is required to be transformed by Hilbert k And (t) converting the spectrum into an analytic signal and then into a single-side spectrum, multiplying the single-side spectrum by a central frequency index term to convert the spectrum of the mode into a corresponding base frequency band, and finally estimating the bandwidth of the demodulation signal through Gaussian smoothing. For photovoltaic power generation, aiming at the relevant meteorological elements which are preliminarily screened in the first step, a key meteorological factor which has a larger influence on the solar energy resources of the place and the photovoltaic power generation capacity of the building is identified by utilizing a multi-element adaptive regression spline method MARS. The MARS divides the data space into a plurality of subareas, adopts a 'break-by-break' strategy, each area has own basis function, and then gradually accumulates a function formula approaching to nonlinearity, so that the data can be better fitted. During the first stage of MARS operation, the MARS constructs a large number of basis functions based on all input interpretation variables; the second stage, through generalized cross The fork validates the GCV to delete those basis functions that do not contribute to the prediction, i.e., to delete those useless interpretation variables. The historical data of the photovoltaic power generation capacity can be used as an interpreted variable, all factors which possibly influence the photovoltaic are used as interpreted variables and input into the MARS, and the factors which can effectively predict the power generation capacity are screened to be used as the input of a follow-up GA-LSTM model through a GCV mechanism of the MARS.
Step 2.1: and carrying out abnormal value detection and correction processing on the data before prediction on the historical wind power generation amount and the historical photovoltaic power generation amount. The detection process by the quartile method comprises the following steps: all data are arranged according to the order of magnitude, and the quartile value is marked as Q 1 I.e. only 1/4 of all data is greater than Q 1 The lower quartile value is Q 2 I.e. only 1/4 of all data is smaller than Q 2 The upper bound is (Q) 1 +1.5(Q 1 -Q 2 ) Lower bound is (Q) 2 -1.5(Q 1 -Q 2 ) Normal observations between the upper and lower bounds, otherwise abnormal values; after the abnormal value is removed, carrying out n times of spline interpolation on abnormal generating capacity data, and setting f (t) as a time interval [ t ] 1 ,t 2 ]The continuous micro-function can remove k abnormal points on the same day, and the generated energy data of 24-k hours are also obtained; let the function f (t) satisfy the condition f (t) =at n +bt n-1 + … +ct+d, in interval [ t ] 1 ,t 2 ]Substituting a normal value in the historical wind power generation data into a second-order continuous derivative, and solving an f (t) expression; the outlier time t is based on the f (t) expression k And f (t) is carried out, and an outlier substitution value is obtained.
Step 2.2: for wind power generation, a nonlinear unstable wind power signal is decomposed into a trend component, a fluctuation component and a random component by utilizing a VMD decomposition method, so that the characteristics of wind power generation are clearly displayed in a layered manner. The trend component may be regarded as a representation of long-term fluctuations of the wind power generation, the fluctuation component may be regarded as a representation of short-term fluctuations, and the random component is close to the random signal. The goal of the VMD is to decompose the real-time input signal f into k subcomponents u with specific sparsity characteristics k . Assuming that the components are compactly centered around their center frequency omega k Frequency ofThe bandwidth in the spectrum determines the a priori sparsity of each component. In the process of decomposing the constructed VMD signal, each modal component u is required to be transformed by Hilbert k And (t) converting the spectrum into an analytic signal and then into a single-side spectrum, multiplying the single-side spectrum by a central frequency index term to convert the spectrum of the mode into a corresponding base frequency band, and finally estimating the bandwidth of the demodulation signal through Gaussian smoothing. The decomposition process aims at minimizing the sum of the estimated bandwidths of all the components, the solved constraint condition is that the sum of all the components is equal to the original input signal, and the specific variational constraint problem expression is as follows:
s.t.∑ k u k =f (3) wherein u k (t) is the kth component after decomposition by a variation mode (here k is specifically 3); omega k Center frequency for each component; delta (t) is a dirac function, specifically f is the original input signal.
To ensure that the accuracy of modal decomposition meets the requirements and that the constraint conditions are strict. Introducing a penalty factor and a Lagrangian multiplier converts the above into an unconstrained optimization problem:
where v is a penalty factor and λ is a lagrangian multiplier. Updating u by iteration k n+1 、ω k n+1 And lambda (lambda) n+1 And obtaining a function optimal solution. Wherein u is k n+1 The corresponding frequency domain can be obtained through Fourier transformation, and the specific function expression is as follows:
step 2.3: for photovoltaic power generation, aiming at related meteorological factors such as sunlight duration, short wave radiation, air temperature, dew point temperature, relative humidity, air pressure and the like which are preliminarily screened in the first step, a key meteorological factor which has great influence on the solar energy resources of the place and the photovoltaic power generation capacity of the building is identified by utilizing a multi-element adaptive regression spline method MARS. Specifically, factors such as sunlight duration, short wave radiation, air temperature, dew point temperature, relative humidity, air pressure and the like are used as the input of the MARS, and building photovoltaic power generation capacity is used as the output of the MARS, so that variables affecting photovoltaic power generation are screened and identified. The MARS divides the data space into a plurality of subareas, adopts a 'break-by-break' strategy, each area has own basis function, and then gradually accumulates a function formula approaching to nonlinearity, so that the data can be better fitted. The specific equation for constructing the MARS recognition model is as follows:
Wherein,represents a basis function, x is an explanatory variable, namely a meteorological element affecting photovoltaic power generation, alpha 0 Is the intercept, a m Is the coefficient of the basis function, M is the number of the basis functions, K m Is the number of nodes, t km Is the value of a node, s km ∈{-1,+1}。
Each basis function exists in two possible forms:
or alternatively
Wherein a is a constant term, namely a demarcation node of different areas.
The main operation of the MARS process is mainly divided into two stages. In the first stage, the MARS constructs a number of basis functions based on all of the input interpretation variables. In the second stage, those basis functions that do not contribute to the prediction, i.e. those useless explanatory variables, are deleted by generalized cross-validation GCV. The definition of GCV is as follows:
where ζ=m+ 1+p (M/2), p is the penalty factor, p e [2,4].
Therefore, the historical data of the photovoltaic power generation capacity can be used as an interpreted variable, all factors which possibly influence the photovoltaic are used as interpreted variables and input into the MARS, and the short-wave radiation, the air temperature and the relative humidity are screened out through a GCV mechanism of the MARS to be used as factors which can effectively predict the power generation capacity and are used as the input of a subsequent GA-LSTM model.
Step three: and constructing a prediction model of the renewable energy power generation amount of the green building. The prediction of wind power generation and photovoltaic power generation is based on the basic framework of the LSTM model for unfolding modeling. LSTM is used as a deep learning network, is specially used for learning time series signals, can store valuable information in memory neurons through an input gate in the learning process, and has better recognition and prediction capability on samples with longer distance while retaining long-term memory. Aiming at an ILSTM prediction model of wind power generation design, the improvement is mainly divided into three parts, namely an input vector, a storage unit and an output unit. The trend component, the fluctuation component and the random component obtained by the VMD decomposition method in the second step are required to be used as model input, and the wind power generation amount is required to be used as output; the random component in the input vector needs to be included as part of the wind power in a short time, but in the long term, due to the unpredictability of the truly random behavior, when the random amount included in the wind power signal is large, the prediction error increases accordingly. Thus, in the improved ILSTM model, the long-term memory of the random component is suppressed, while the short-term memory of the random component is maintained. And (3) based on the principle of a sliding window, forming a sample comprising a plurality of time dimensions by using a trend component, a fluctuation component and a random component obtained by the VMD decomposition method in the step two as an input vector, wherein the output of the prediction model is the predicted value of the wind power generation amount. The prediction model identifies and learns the modes represented in the input vector, and establishes a mapping relation between the input vector and the output vector, so that the prediction capability of the network model is realized. In order to avoid the influence of random components on the long-term memory mode, a parameter vector D= (1, alpha) is added to the input of the component memory unit, wherein the randomness of the wind power signal is evaluated based on the prediction error to determine the value of the random component long-term memory unit inhibition parameter alpha. And designing a GA-LSTM prediction model aiming at photovoltaic power generation, wherein short wave radiation, air temperature and relative humidity which are key meteorological factors identified by the MARS in the second step are used as input, and photovoltaic power generation is used as output. LSTM has numerous superparameters to evaluate in training and predicting data. The super parameters of the model are optimized by introducing a genetic algorithm into the basic LSTM model, so that the average absolute percentage error MAPE in the prediction process is minimized, and a numerical solution is obtained for the problem based on the genetic algorithm GA-AM of the adaptive variation. Genetic algorithms are algorithms that solve unconstrained and constrained nonlinear optimization problems based on natural selection processes that mimic biological evolution. The algorithm iteratively modifies the population of individual solutions. Firstly, constructing a plurality of super-parameter combinations from a given super-parameter space by a genetic algorithm; then each super-parameter combination is brought into a GA-LSTM model, training and predicting are carried out on the data, and an objective function value MAPE is calculated; selecting a super-parameter combination with smaller MAPE (MAPE) as a parent based on the accounting of the first two steps, and continuously performing cross recombination based on the selected super-parameter combination to generate a child; finally, in a given parameter space, the super-parameter combination is randomly adjusted until the best super-parameter combination is found, and super-parameter tuning of the basic LSTM model is completed.
Step 3.1: wind power generation and photovoltaic power generation prediction are both based on an LSTM model basic framework for unfolding modeling. LSTM is used as a deep learning network, is specially used for learning time series signals, is added with a memory neuron mechanism on the basis of an RNN model, and is more suitable for processing long-term information. Specifically, the LSTM model introduces forgetting gate, input gate and output gate modulation and updates the state of memory neurons.
The first step in the LSTM model is to let the inputs at time t and the outputs of the neurons at time t-1 enter the forget gate, which decides which messages should be removed.
x t Is the input of the moment t, h t-1 Is the output of the neuron at time t-1,and->Is a weight matrix, < >>Is an error vector, sigmoid (z) = (1+e) -z ) -1 Is an activation function.
The second step of the LSTM model is to enter the input at time t and the output of the neuron at time t-1 into the input gate while calculating the candidate value of the neuron state.
Wherein,is a candidate for neuronal status, i t Is an input door, is a->And->Is a weight matrix, < >>And->Is an error vector, +.>
The third step of the LSTM model is to update the state of the neuron by the state of the neuron at the time t-1, the information of the amnestic gate and the input gate screening and the candidate value of the neuron.
Wherein, the ". Is the multiplication of the corresponding elements.
The fourth step of the LSTM model is to output the information output by the gate through the updated neuron state, and generate the output of the neuron at the time t.
h t ＝o t ⊙tanh(c t )
Wherein,and->Is a weight matrix, < >>Is an error vector o t Is an output door h t Is the output of the neuron state.
Step 3.2: for an ILSTM prediction model designed by wind power generation, the improvement is mainly divided into three parts of an input vector, a storage unit and an output unit, and the model structure is shown in figure 2. The trend component, the fluctuation component and the random component obtained by the VMD decomposition method in the second step are required to be used as model input, and the wind power generation amount is required to be used as output; the random component in the input vector needs to be included as part of the wind power in a short time, but in the long term, due to the unpredictability of the truly random behavior, when the random amount included in the wind power signal is large, the prediction error increases accordingly. Thus, in the improved ILSTM model, the long-term memory of the random component is suppressed, while the short-term memory of the random component is maintained.
Based on the principle of a sliding window, the trend component, the fluctuation component and the random component obtained by the VMD decomposition method in the second step are formed into a sample comprising a plurality of time dimensions as input vectors, and the specific form is as follows:
x kt ＝(u k (t-n+1),u k (t-n+2),…,u k (t))
Wherein x is kt The input vector at the time of the kth component t after VMD decomposition is given, and n is the sliding time window width.
And the output of the prediction model is the predicted value of the wind power generation amount. The prediction model identifies and learns the modes represented in the input vector, and establishes a mapping relation between the input vector and the output vector, so that the prediction capability of the network model is realized. To avoid the effect of random components on the long-term memory pattern, the parameter D is added to the input of the component memory cell:
wherein, D is the suppression vector parameter of the whole three components, D= (1, alpha), alpha is the specific random component long-term memory unit suppression parameter, and the alpha is more than or equal to 0 and less than or equal to 1. The specific meaning of this part of parameters is as follows:
d is designed to be (1, a) so that the first two-tier components can be transferred directly into long-term memory, and the third tier components multiplied by a before entering the memory cell. Because 0.ltoreq.α.ltoreq.1, the value of the random component is suppressed. The value of alpha depends on the content of random components in the wind power signal. When the randomness is strong, the value of α should be small. For example, if a is designed to be 0, the third layer component with random components will be completely eliminated and not included in long-term memory. When the wind power signal randomness is weak, the value of alpha can be designed to be 1. If α is 1, the third layer component will go entirely to long term memory. Evaluating the randomness of the wind power signal based on the prediction error to determine the value of alpha:
α＝1-err*10
Wherein,
in the ILSTM, the output gate and output are defined as follows:
o′ t ＝W ox x t +b o
h′ t ＝o′ t ⊙φ(c′ t )
step 3.3: and designing a GA-LSTM prediction model aiming at photovoltaic power generation, wherein short wave radiation, air temperature and relative humidity which are key meteorological factors identified by the MARS in the second step are used as inputs, photovoltaic power generation capacity is used as output, and the model structure is shown in figure 3. LSTM has numerous superparameters to evaluate in training and predicting data. The super parameters of the model are optimized by introducing a genetic algorithm into the basic LSTM model, and the following objective function is set for minimizing the average absolute percentage error MAPE in the prediction process:
wherein y is t Is the actual value of the current,is the predicted value, T is the time length, H is the number of hidden units, E is the training number, B is the data batch size, R is the learning rate, F is the learning rate decline factor, P is the learning rate decline period, G is the gradient threshold, Ω is the supersoundParameter space.
Since it is difficult to solve the constrained optimization problem by means of calculus, a genetic algorithm GA-AM based on adaptive mutation is used to solve the problem by means of numerical solution, and the model structure is shown in fig. 4. Genetic algorithms are algorithms that solve unconstrained and constrained nonlinear optimization problems based on natural selection processes that mimic biological evolution. The algorithm iteratively modifies the population of individual solutions. At each step, the genetic algorithm randomly selects individuals from the current population and uses them as parents to generate next generation offspring. After one generation and another, the population "evolves" into an optimal solution.
The first step of the genetic algorithm is to construct a plurality of superparameter combinations from within a given superparameter space.
The second step of the genetic algorithm is to bring each super-parameter combination into the GA-LSTM model, train and predict the data, and calculate the objective function value, i.e. MAPE.
The third step of the genetic algorithm is to select two sets of better hyper-parameter combinations, i.e. combinations that make MAPE smaller, as parents.
The fourth step of the genetic algorithm is to perform cross-recombination based on the selected hyper-parametric combinations to produce offspring.
And a fifth step of genetic algorithm, wherein the super-parameter combination is randomly adjusted in a given parameter space.
The genetic algorithm continuously repeats the process until the best super-parameter combination is found, and the super-parameter tuning of the basic LSTM model is completed.
Step four: and training a green building renewable energy generating capacity prediction model.
Step 4.1: normalizing historical data of wind power trend components, fluctuation components and random components which are obtained by a VMD decomposition method in the second step and weather factors which influence photovoltaic power generation and are identified by a MARS regression method; normalized data were processed at 6:2: the proportion of 2 is divided into a training set, a verification set and a test set;
Step 4.2: the rolling window is used for training and predicting the data, and the specific structure is shown in fig. 5. Inputting the three wind component signals and the historical wind power generation data of the corresponding time period into the ILSTM model built in the third step to train a wind power generation prediction model, wherein the time step is set to be 3 hours; and (3) inputting relevant historical meteorological data and historical photovoltaic power generation data of a corresponding period into the GA-LSTM model training photovoltaic power generation prediction model built in the third step, wherein the time step is set to be 3 hours.
Step five: and outputting a prediction result of the renewable energy power generation amount of the green building according to the trained renewable energy power generation amount prediction model of the green building.
Step 5.1: acquiring historical wind power generation amount data of the first 1 day and the first 1 week of prediction, decomposing a wind power signal into a trend component, a fluctuation component and a random component according to the method of the second step, acquiring a forecast value of a key meteorological factor which influences photovoltaic power generation amount on the current day and the current week of prediction, and carrying out normalization processing on related data of wind power generation and photovoltaic power generation amount;
step 5.2: inputting the processed VMD decomposition components into a trained ILSTM model, and respectively outputting wind power generation capacity prediction results of 1 day in the short term and 1 week in the middle term; inputting the processed critical weather factor forecast values into a trained GA-LSTM model, and respectively outputting photovoltaic power generation quantity forecast results of 1 day in the short term and 1 week in the medium term;
Step 5.3: and adding the predicted wind power generation amount and photovoltaic power generation amount data to obtain a predicted result of the short-term power generation amount in the renewable energy sources of the green building.
The present embodiment will explain the effects of the present invention in two ways.
Experimental conditions
The embodiment discloses a prediction method of renewable energy power generation capacity of a green building, which takes a small green building sitting in Beijing as a design example. The building area of the house is 36 meters, the length and the width are 6 meters, the height is 2.5 meters, 2 500W wind driven generators and 8 450W monocrystalline silicon solar photovoltaic panels are mounted on the roof to supply green electricity, and wind power and photovoltaic solar energy can generate electricity by 2 degrees and 11 degrees. Meanwhile, the cabin is provided with 24 100AH lithium batteries, and can store 28-DEG electricity at maximum. In the modeling analysis process of the experiment, training and tracking are carried out by using a MATLAB deep learning tool box under the environment of Intel i7-1360P CPU 5 GHz.
(II) results of experiments
The example "green house" achieves good predictive results for the short before day and the mid before week over a 50 day test period. The solar prediction precision of photovoltaic and wind power generation in a short period reaches 93 percent and 85 percent respectively, and the solar prediction effective rate of the whole renewable energy source power generation in the cabin exceeds 90 percent. The prediction is performed in the middle period for 7 days, the prediction effective rate of the whole power generation is 80%, and the energy supply prediction accuracy of the energy self-contained small building is effectively improved.
While the foregoing is directed to embodiments of the present invention, other and further embodiments of the invention may be devised without departing from the basic scope thereof, and the scope thereof is determined by the claims that follow.
Claims (7)
1. A renewable energy prediction method for a green building is characterized in that: comprises the following steps of the method,
step one: collecting historical data of wind and light generating capacity of a green building and relevant meteorological factors influencing photovoltaic generating capacity; respectively acquiring high-frequency wind power generation capacity historical data and photovoltaic power generation capacity historical data from a data acquisition system of a green building wind power inverter and a photovoltaic inverter; from a photovoltaic power generation mechanism, potential meteorological elements influencing photovoltaic power generation are determined, and through checking with a preset authoritative meteorological product type, the selected meteorological indexes can be ensured to be capable of acquiring historical observation data of an hour level, and weather forecast data of a county meteorological station where a building is located can be provided by a weather forecast system to serve as input variables for the prediction of the follow-up photovoltaic power generation; setting a matching area taking a building position as a center and a preset distance as a radius on the selection of meteorological stations, and performing inverse distance weighted interpolation on the values of all the meteorological stations in the area;
Step two: after finishing the data preprocessing according to the first step, carrying out feature selection on the wind power generation amount data, and carrying out factor identification on the photovoltaic power generation amount data; carrying out abnormal value detection and correction processing on the data before prediction on the historical wind power generation amount and the historical photovoltaic power generation amount, detecting abnormal values through a quartile method, and carrying out n times of spline interpolation processing on abnormal power generation amount data to obtain abnormal value substitution points; for wind power generation, decomposing a nonlinear unstable wind power signal into a trend component, a fluctuation component and a random component by using a VMD decomposition method; the trend component is equivalent to the representation of long-term fluctuation of wind power generation, the fluctuation component is equivalent to the representation of short-term fluctuation, and the random component is close to a random signal; during the process of decomposing the constructed VMD signal, each modal component u is transformed by Hilbert k (t) converting the spectrum into an analytic signal and then into a single-side spectrum, multiplying the single-side spectrum by a central frequency index term to convert the modal spectrum into a corresponding base frequency band, and estimating the bandwidth of the demodulation signal through Gaussian smoothing; for photovoltaic power generation, aiming at the relevant meteorological elements preliminarily screened in the first step, identifying key meteorological factors which have great influence on the solar energy resources of the location and the photovoltaic power generation capacity of the building by utilizing a multi-element adaptive regression spline method MARS; dividing the data space into a plurality of subareas by the MARS, adopting a 'break-by-break' strategy, wherein each area has own basis function, and gradually accumulating a function formula approaching to nonlinearity, thereby improving the accuracy and efficiency of data fitting; during the MARS operation, in the first stage, the MARS constructs a number of basis functions based on all input interpretation variables; a second stage of deleting those basis functions that do not contribute to the prediction, i.e. deleting those useless explanatory variables, by generalized cross-validation GCV; taking historical data of photovoltaic power generation capacity as an interpreted variable, taking all factors possibly influencing the photovoltaic as interpreted variables, inputting the interpreted variables into a MARS, and screening the factors capable of effectively predicting the power generation capacity as the input of a subsequent GA-LSTM model through a GCV mechanism of the MARS;
Step three: constructing a prediction model of renewable energy power generation of a green building; the prediction of wind power generation and photovoltaic power generation is based on the basic framework of an LSTM model for unfolding modeling; aiming at an ILSTM prediction model designed by wind power generation, the improvement is mainly divided into three parts, namely an input vector, a storage unit and an output unit; taking a trend component, a fluctuation component and a random component which are obtained by the VMD decomposition method in the second step as model input and taking wind power generation amount as output; the random component in the input vector needs to be included as part of the wind power in a short time; based on the principle of a sliding window, the trend component, the fluctuation component and the random component obtained by the VMD decomposition method in the second step are formed into a sample comprising a plurality of time dimensions to be used as input vectors, and the output of the prediction model is the predicted value of the wind power generation capacity; the prediction model identifies and learns the mode represented in the input vector, and establishes a mapping relation between the input vector and the output vector; adding a parameter vector D= (1, alpha) into the input of the component memory unit, wherein the randomness of the wind power signal is evaluated based on the prediction error, and the value of the inhibition parameter alpha of the random component long-term memory unit is determined; designing a GA-LSTM prediction model aiming at photovoltaic power generation, wherein key meteorological factors which influence photovoltaic power generation capacity and are identified by a MARS in the second step are taken as input, and photovoltaic power generation capacity is taken as output; introducing a genetic algorithm into a basic LSTM model to perform optimization on the super parameters of the model, so that the average absolute percentage error MAPE in the prediction process is minimum, and solving the problem by using a genetic algorithm GA-AM based on adaptive variation; constructing a plurality of super-parameter combinations from a given super-parameter space based on a genetic algorithm; then each super-parameter combination is brought into a GA-LSTM model, training and predicting are carried out on the data, and an objective function value MAPE is calculated; selecting a super-parameter combination with smaller MAPE (MAPE) as a parent based on the accounting of the first two steps, and continuously performing cross recombination based on the selected super-parameter combination to generate a child; in a given parameter space, the super-parameter combination is randomly adjusted until the best super-parameter combination is found, and super-parameter tuning of the basic LSTM model is completed;
Step four: normalizing historical data of wind power trend components, fluctuation components and random components which are obtained by a VMD decomposition method in the second step and weather factors which influence photovoltaic power generation and are identified by a MARS regression method; dividing the normalized data into a training set, a verification set and a test set; inputting the three wind component signals and the historical wind power generation data of the corresponding time period into the ILSTM model built in the third step to train a wind power generation prediction model; inputting relevant historical meteorological data and historical photovoltaic power generation data of a corresponding period into the GA-LSTM model built in the third step to train a photovoltaic power generation prediction model, and obtaining a trained green building renewable energy power generation prediction model;
step five: acquiring historical wind power generation data of a predicted preset time period, and decomposing a wind power signal into a trend component, a fluctuation component and a random component according to the second step; acquiring forecast values of key meteorological factors which influence photovoltaic power generation on the current day and the current week, and carrying out normalization processing on related data of the wind power generation and the photovoltaic power generation; inputting the processed VMD decomposition components into a trained ILSTM model, and respectively outputting wind power generation capacity prediction results corresponding to a preset time period; inputting the processed critical weather factor forecast values into a trained GA-LSTM model, and respectively outputting photovoltaic power generation quantity forecast results corresponding to a preset time period; and adding the predicted wind power generation amount and photovoltaic power generation amount data to obtain a predicted result of the short-term power generation amount in the renewable energy sources of the green building.
2. The method for predicting renewable energy sources of green buildings according to claim 1, wherein: the first implementation method of the step is that,
step 1.1: acquiring high-frequency wind power generation capacity historical data and photovoltaic power generation capacity historical data, and ensuring that the sample size meets the preset requirement of the follow-up neural network model prediction;
step 1.2: from a photovoltaic power generation mechanism, determining potential meteorological elements influencing photovoltaic power generation capacity, including variables such as sunlight duration, short wave radiation, air temperature, dew point temperature, relative humidity, air pressure and the like, and acquiring observation and forecast data of a county meteorological site where a building is located;
step 1.3: on the selection of meteorological stations, setting a matching area taking a building position as a center and a preset distance as a radius, and carrying out inverse distance weighted interpolation on the numerical values of all the meteorological stations in the area, wherein a calculation formula of a specific inverse distance weighted interpolation method is as follows:
wherein AD (x, y) is the atmospheric figure value of the green building site obtained by final interpolation, AD is the meteorological observation value of the site in the ith matching zone, d i For the distance between the i-th known point and the interpolation point, p is the interpolation parameter.
3. The method for predicting renewable energy sources of green buildings according to claim 2, wherein: the implementation method of the second step is that,
Step 2.1: detecting abnormal values of data by using a quartile method before prediction for the historical wind power generation amount and the historical photovoltaic power generation amount, and correcting the vacant power generation amount data by using n times of spline interpolation after the abnormal values are removed;
step 2.2: for wind power generation, a VMD decomposition method is utilized to decompose a nonlinear unstable wind power signal into 3 components of a trend component, a fluctuation component and a random component, so that characteristics of long-term fluctuation, short-term fluctuation and random signals of the wind power generation are clearly displayed in a layered manner; during the process of decomposing the constructed VMD signal, each modal component u is transformed by Hilbert k (t) converting the spectrum into an analytic signal and then into a single-side spectrum, multiplying the single-side spectrum by a central frequency index term to convert the modal spectrum into a corresponding base frequency band, and estimating the bandwidth of the demodulation signal through Gaussian smoothing; the decomposition process aims at minimizing the sum of the estimated bandwidths of all the components, the solved constraint condition is that the sum of all the components is equal to the original input signal, and the specific variational constraint problem expression is as follows:
s.t.∑ k u k ＝f (3)
wherein u is k (t) is the kth component after decomposition by the variational mode; omega k Center frequency for each component; delta (t) is a dirac function, specifically f is the original input signal;
to ensure that the accuracy of modal decomposition meets the requirements and the strictness of constraint conditions is ensured; introducing a penalty factor and a Lagrangian multiplier converts the above into an unconstrained optimization problem:
wherein v is a penalty factor and lambda is a Lagrangian multiplier; updating u by iteration k n+1 、ω k n+1 And lambda (lambda) n+1 Obtaining a function optimal solution; wherein u is k n+1 The corresponding frequency domain is obtained through Fourier transformation, and the specific function expression is as follows:
step 2.3: for photovoltaic power generation, aiming at relevant meteorological elements which are preliminarily screened in the first step, identifying key meteorological factors which have great influence on local solar resources and building photovoltaic power generation capacity by utilizing a multi-element adaptive regression spline method MARS, wherein the relevant meteorological elements comprise sunlight duration, short wave radiation, air temperature, dew point temperature, relative humidity and air pressure which are used as inputs of the MARS, and building photovoltaic power generation capacity which is used as outputs of the MARS, so that variables which influence photovoltaic power generation are screened and identified; the specific equation for constructing the MARS recognition model is as follows:
wherein,represents a basis function, x is an explanatory variable, namely a meteorological element affecting photovoltaic power generation, alpha 0 Is the intercept, a m Is the coefficient of the basis function, M is the number of the basis functions, K m Is the number of nodes, t km Is the value of a node, s km ∈{-1,+1}；
Each basis function exists in two possible forms:
or alternatively
A is a constant term, namely a demarcation node of different areas;
the main operation process of the MARS method is mainly divided into two stages; the first stage, the MARS constructs a number of basis functions based on all the input interpretation variables; a second stage of deleting those basis functions that do not contribute to the prediction, i.e. deleting those useless explanatory variables, by generalized cross-validation GCV; the definition of GCV is as follows:
where ζ=m+ 1+p (M/2), p is the penalty factor, p e [2,4];
therefore, the historical data of the photovoltaic power generation capacity is taken as an interpreted variable, all factors which possibly influence the photovoltaic are taken as interpreted variables and input into the MARS, and the factors which can effectively predict the power generation capacity are screened to be taken as the input of a subsequent GA-LSTM model through a GCV mechanism of the MARS.
4. A method of renewable energy prediction for green buildings as defined in claim 3, wherein: the implementation method of the third step is that,
step 3.1: the prediction of wind power generation and photovoltaic power generation is based on the basic framework of an LSTM model, and the LSTM model of the memory neuron introduces forgetting gate, input gate and output gate regulation and updates the state of the memory neuron;
The first step of LSTM model is to make the input at time t and the output of neuron at time t-1 enter into forget gate, which decides which messages should be removed;
x t is the input of the moment t, h t-1 Is the output of the neuron at time t-1,and->Is a weight matrix, < >>Is an error vector, sigmoid (z) = (1+e) -z ) -1 Is an activation function;
the second step of the LSTM model is to enable the input at the moment t and the output of the neuron at the moment t-1 to enter an input gate, and simultaneously calculate candidate values of the state of the neuron;
wherein,is a candidate for neuronal status, i t Is an input door, is a->And->Is a weight matrix, < >>And->Is an error vector, +.>
The third step of LSTM model is to update the state of the neuron by the state of the neuron at time t-1, the information screened by the forgetting gate and the input gate, and the candidate value of the neuron;
wherein, the ";
outputting the information output by the gate through the updated neuron state to generate the output of the neuron at the moment t;
h t ＝o t ⊙tanh(c t ) (15)
wherein,and->Is a weight matrix, < >>Is an error vector o t Is an output door h t Is the output of the neuron state;
step 3.2: aiming at an ILSTM prediction model designed by wind power generation, the improvement is mainly divided into three parts, namely an input vector, a storage unit and an output unit; the trend component, the fluctuation component and the random component obtained by the VMD decomposition method in the second step are required to be used as model input, and the wind power generation amount is required to be used as output; the random component in the input vector needs to be taken into the input vector as a part of wind power in a short time, but the prediction error is increased correspondingly when the random quantity contained in the wind power signal is large in the long term due to the fact that the real random behavior is unpredictable; thus, in the improved ILSTM model, the long-term memory of the random component is suppressed, while the short-term memory of the random component is maintained;
Step 3.3: designing a GA-LSTM prediction model aiming at photovoltaic power generation, wherein key meteorological factors which influence photovoltaic power generation capacity and are identified by a MARS in the second step are taken as input, and photovoltaic power generation capacity is taken as output; in the training and predicting of data, LSTM has a plurality of super parameters to be estimated; and optimizing the super parameters of the model by introducing a genetic algorithm into the basic LSTM model.
5. The method for predicting renewable energy sources of green buildings according to claim 4, wherein the method comprises the following steps: the specific method of the step 3.2 is that,
based on the principle of a sliding window, the trend component, the fluctuation component and the random component obtained by the VMD decomposition method in the second step are formed into a sample comprising a plurality of time dimensions as input vectors, and the specific form is as follows:
x kt ＝(u k (t-n+1),u k (t-n+2),…,u k (t)) (16)
wherein x is kt The input vector of the kth component t moment after VMD decomposition is given, and n is the width of the sliding time window;
the output of the prediction model is the predicted value of the wind power generation amount; the prediction model identifies and learns the mode represented in the input vector, and establishes a mapping relation between the input vector and the output vector, so that the prediction capability of the network model is realized; to avoid the effect of random components on the long-term memory pattern, the parameter D is added to the input of the component memory cell:
Wherein, D is the suppression vector parameter of the whole three components, D= (1, alpha), alpha is the specific random component long-term memory unit suppression parameter, and the alpha is more than or equal to 0 and less than or equal to 1; the specific meaning of this part of parameters is as follows:
d is designed as (1, α) so that the first two-tier component can be transferred directly into long-term memory, the third tier component being multiplied by α before entering the memory cell; because 0.ltoreq.α.ltoreq.1, the value of the random component is suppressed; the value of alpha depends on the content of random components in the wind power signal; when the randomness is strong, the value of α should be small; for example, if α is designed to be 0, the third layer component with random component will be completely eliminated and not included in long-term memory; when the randomness of the wind power signal is weak, the value of alpha can be designed to be 1; if α is 1, the third layer component will fully enter long term memory; evaluating the randomness of the wind power signal based on the prediction error to determine the value of alpha:
α＝1-err*10 (18)
wherein,
in the ILSTM, the output gate and output are defined as follows:
o′ t ＝W ox x t +b o (19)
h′ t ＝o′ t ⊙φ(c′ t ) (20) 。
6. the method for predicting renewable energy sources of green buildings according to claim 4, wherein the method comprises the following steps: the specific method of the step 3.3 is as follows:
taking the key meteorological factors which are identified by the MARS and influence the photovoltaic power generation amount in the second step as input and taking the photovoltaic power generation amount as output;
Introducing a genetic algorithm into a basic LSTM model to perform optimization on the super parameters of the model, and setting the following objective function to minimize the average absolute percentage error MAPE in the prediction process:
wherein y is t Is the actual value of the current,is a predicted value, T is a time length, H is the number of hidden units, E is the training times, B is the data batch size, R is the learning rate, F is the learning rate reduction factor, P is the learning rate reduction period, G is a gradient threshold, and Ω is a hyper-parameter space;
because the constrained optimization problem is difficult to solve by means of calculus, a genetic algorithm GA-AM based on adaptive variation is used for solving the problem by means of solving a numerical solution;
the first step of the genetic algorithm is to construct a plurality of super-parameter combinations from the given super-parameter space;
the second step of the genetic algorithm is to bring each super-parameter combination into a GA-LSTM model, train and predict data, and calculate objective function values, namely MAPE;
the third step of the genetic algorithm is to select two sets of better hyper-parameter combinations, namely combinations with smaller MAPE as parents;
the fourth step of genetic algorithm is to carry out cross recombination based on the selected hyper-parameter combination to generate offspring;
The fifth step of genetic algorithm is to randomly adjust the super parameter combination in the established parameter space;
the genetic algorithm continuously repeats the process until the best super-parameter combination is found, and the super-parameter tuning of the basic LSTM model is completed.
7. The method for predicting renewable energy sources of green buildings according to claim 6, wherein the method comprises the following steps: step five, acquiring historical wind power generation capacity data of a predicted preset time period, and selecting the historical wind power generation capacity data of 1 day and 1 week before the prediction;
inputting the processed VMD decomposition components into a trained ILSTM model, respectively outputting wind power generation quantity prediction results corresponding to a preset time period, and selecting the wind power generation quantity prediction results to be respectively output for 1 day in the short term and 1 week in the middle term;
and inputting the processed key weather factor forecast values into a trained GA-LSTM model, respectively outputting photovoltaic power generation quantity forecast results corresponding to a preset time period, and selecting the photovoltaic power generation quantity forecast results to be respectively output for 1 day in the short term and 1 week in the medium term.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311522411.2A CN117728386A (en) | 2023-11-15 | 2023-11-15 | Prediction method for renewable energy power generation capacity of green building |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311522411.2A CN117728386A (en) | 2023-11-15 | 2023-11-15 | Prediction method for renewable energy power generation capacity of green building |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117728386A true CN117728386A (en) | 2024-03-19 |
Family
ID=90198743
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311522411.2A Pending CN117728386A (en) | 2023-11-15 | 2023-11-15 | Prediction method for renewable energy power generation capacity of green building |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117728386A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118297359A (en) * | 2024-06-05 | 2024-07-05 | 华维设计集团股份有限公司 | Green energy-saving design optimization method for building engineering |
-
2023
- 2023-11-15 CN CN202311522411.2A patent/CN117728386A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118297359A (en) * | 2024-06-05 | 2024-07-05 | 华维设计集团股份有限公司 | Green energy-saving design optimization method for building engineering |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Kaloop et al. | Novel application of adaptive swarm intelligence techniques coupled with adaptive network-based fuzzy inference system in predicting photovoltaic power | |
| CN112348271A (en) | Short-term photovoltaic power prediction method based on VMD-IPSO-GRU | |
| Mellit et al. | An ANFIS-based forecasting for solar radiation data from sunshine duration and ambient temperature | |
| Eom et al. | Feature-selective ensemble learning-based long-term regional PV generation forecasting | |
| Natarajan et al. | Forecasting energy generation in large photovoltaic plants using radial belief neural network | |
| Kalogirou et al. | Artificial intelligence techniques in solar energy applications | |
| Niu et al. | Uncertainty modeling for chaotic time series based on optimal multi-input multi-output architecture: Application to offshore wind speed | |
| CN110942205B (en) | Short-term photovoltaic power generation power prediction method based on HIMVO-SVM | |
| Yin et al. | A novel asexual-reproduction evolutionary neural network for wind power prediction based on generative adversarial networks | |
| CN112100911B (en) | Solar radiation prediction method based on depth BILSTM | |
| CN112434848B (en) | Nonlinear weighted combination wind power prediction method based on deep belief network | |
| CN112215428B (en) | Photovoltaic power generation power prediction method and system based on error correction and fuzzy logic | |
| CN110648017A (en) | Short-term impact load prediction method based on two-layer decomposition technology | |
| Zhang et al. | Wind speed prediction research considering wind speed ramp and residual distribution | |
| CN117728386A (en) | Prediction method for renewable energy power generation capacity of green building | |
| CN115496263A (en) | Short-term photovoltaic power generation prediction method based on space-time genetic attention network | |
| Souabi et al. | Data-driven prediction models of photovoltaic energy for smart grid applications | |
| CN116432812A (en) | New energy power prediction method for optimizing LSTM (least squares) by using Zun sea squirt algorithm | |
| CN118095891A (en) | Active power distribution network payload prediction method and system considering source load meteorological characteristic decoupling | |
| Phan et al. | Application of a new Transformer-based model and XGBoost to improve one-day-ahead solar power forecasts | |
| Tahir et al. | Enhanced solar photovoltaic power prediction using diverse machine learning algorithms with hyperparameter optimization | |
| CN113112085A (en) | New energy station power generation load prediction method based on BP neural network | |
| Zhang et al. | Deep reinforcement learning based interpretable photovoltaic power prediction framework | |
| Boubaker et al. | Prediction of Daily Global Solar Radiation using Resilient-propagation Artificial Neural Network and Historical Data: A Case Study of Hail, Saudi Arabia. | |
| Sim et al. | Temporal Attention Gate Network With Temporal Decomposition for Improved Prediction Accuracy of Univariate Time-Series Data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |