CN115860177A - Photovoltaic power generation power prediction method based on combined machine learning model and application thereof - Google Patents
Photovoltaic power generation power prediction method based on combined machine learning model and application thereof Download PDFInfo
- Publication number
- CN115860177A CN115860177A CN202211325112.5A CN202211325112A CN115860177A CN 115860177 A CN115860177 A CN 115860177A CN 202211325112 A CN202211325112 A CN 202211325112A CN 115860177 A CN115860177 A CN 115860177A
- Authority
- CN
- China
- Prior art keywords
- data
- photovoltaic
- prediction
- item
- equal
- 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
Images
Abstract
The photovoltaic power generation power prediction method based on the combined machine learning model and the application thereof comprise the following steps: preprocessing photovoltaic output data; analyzing the characteristics of photovoltaic output data, and extracting intrinsic mode vectors and residual terms of the power data of the photovoltaic system by adopting an Empirical Mode Decomposition (EMD) algorithm; calculating sampling entropy SE of each eigenmode vector of the photovoltaic output data; distinguishing a trend item and a fluctuation item of the photovoltaic data, setting a threshold value by analyzing the sampling entropy of the whole photovoltaic output sequence, and superposing the eigenvector sampling entropies larger than the threshold value as the fluctuation item and superposing the eigenvector sampling entropies smaller than the threshold value as the trend item; predicting the trend item by adopting a long-short term memory (LSTM) method; and predicting the fluctuation items by adopting a Support Vector Regression (SVR) method to construct a digital twin model for photovoltaic power generation power prediction, and calculating and updating the superposition weight of the prediction data of the photovoltaic trend items and the fluctuation items in real time by using a PSO algorithm in combination with the actual power data of the photovoltaic system.
Description
Technical Field
The invention relates to a prediction method and application thereof, in particular to a photovoltaic power generation power prediction method based on a combined machine learning model and application thereof. Photovoltaic power prediction is realized through an empirical mode decomposition method, long-term and short-term memory and a support vector regression algorithm; based on a particle swarm algorithm, the prediction model is corrected by using the output data of the actual photovoltaic system, so that the digital twin of photovoltaic power generation power prediction is realized, and the method belongs to the technical field of intelligent power.
Background
Modeling and prediction of output power in a photovoltaic system are crucial to the safe and stable operation of a novel power system mainly based on new energy. The high-precision photovoltaic system power prediction can reduce the influence of photovoltaic power generation uncertainty on a power grid and improve the power quality and the penetration level of a photovoltaic system. The photovoltaic power generation has high volatility and intermittency, and the high-precision modeling and prediction of the output power of the photovoltaic system can effectively improve the effective operation of the energy management system and the grid-connected dispatching management level of the power system.
The photovoltaic power generation power prediction mainly adopts a physical method, a statistical method and an artificial intelligence method. The prediction is realized based on a physical method, a large number of sensors are needed to obtain physical parameters, so that mechanism modeling is carried out, but the robustness of the prediction method is poor. The statistical method and the artificial intelligence method are based on data-driven modeling, the traditional statistical method maps the relation between the generated energy and the historical data through methods such as regression analysis and the like, but the effect of processing nonlinear data is not ideal, and compared with the artificial intelligence technology, the artificial intelligence technology has the capability of processing complex nonlinear problems and stronger fault tolerance.
In the mechanism-based modeling method, due to the fact that uncertain factors of prediction results obtained by physical method calculation are too many, the modeling process is complex, and some abnormal weather conditions are difficult to simulate, besides, parameters of a photovoltaic module slowly change along with time, the model is poor in anti-interference capability, low in robustness, large in calculation result deviation and not beneficial to scheduling and planning work of a subsequent system, and the existing prediction technology mainly focuses on a statistical method and an artificial intelligence method.
Statistical methods predictive methods include time series methods, linear regression methods, and time trend extrapolation methods. In the time series method, the two most used time series models are the autoregressive moving average model (ARMA) and the differential integrated moving average autoregressive model (ARIMA). The ARMA model is a time series model consisting of a "mixture" of Autoregressive (AR) and Moving Average (MA) models. The ARIMA model is an extended ARMA model, the ARIMA mainly corresponds to a non-stationary time sequence, the ARMA corresponds to a stationary time sequence, and the time sequence method model is simple but has low overall prediction capability. The linear regression method utilizes a linear regression equation to model the photovoltaic power generation and the independent variables related to the photovoltaic power generation, so as to predict the power generation. At present, a multiple linear regression method is more applied. For example, a new irradiance slant conversion method is proposed in the document [1], and numerical weather forecast horizontal irradiance is converted into numerical weather forecast slant irradiance, so that a short-term photovoltaic prediction statistical method is performed. However, the regression method has a small amount of calculation, but the prediction accuracy is not high. The time trend extrapolation method is used for extrapolating future photovoltaic power generation change according to a historical time sequence of the photovoltaic power generation and mainly comprises a gray prediction model and a Markov chain model, wherein the gray prediction model is generated by accumulating an original sequence once substantially, so that the generated sequence is in a certain rule, and a typical curve is used for fitting. For example, in the document [2], spatial correlation analysis is performed on a photovoltaic power station by utilizing gray correlation analysis, then the photovoltaic power station highly similar to an unknown power station is selected, and the time-space characteristics of the unknown power station are obtained based on a time sequence prediction algorithm GeoMAN model, so that the photovoltaic power generation prediction of the unknown power station with higher precision is realized. The Markov chain model mainly researches the states and state transitions of things. In the photovoltaic power generation prediction, a photovoltaic system is regarded as a whole by the model, firstly, a state vector of photovoltaic power generation is formed according to the change trends of solar radiation amount and environment temperature at different moments every day, and then the power generation of a future system is determined according to a transition probability matrix of the power generation at each moment. For example, in document [3], after the weather state historical samples are classified and the solar radiation model is corrected, a multi-order weighted Markov chain based on an error sequence is established to output an irradiance predicted value. However, in general, the above mathematical statistics model is relatively simple, but due to the general prediction accuracy, the research is not much. For example, in document [4], the effect of an artificial neural network (bidirectional long-term short-term memory) and a statistical method (seasonal autoregressive integrated moving average) model (such as ARMA, ARIMA and SARIMA) on prediction of photovoltaic power generation is compared, and a comparison study shows that the neural network is more accurate than an implemented statistical model and has less calculation time when photovoltaic power generation prediction based on a time series is carried out.
The artificial intelligence prediction algorithm is a popular technology for power generation prediction in the existing photovoltaic system. Long-short term memory (LSTM) is currently widely used for prediction of photovoltaic power generation. The LSTM comprises an input gate, a forgetting gate, an output gate and an internal memory unit, wherein the internal memory unit, a new memory internal unit and an output vector calculation value are selected by the gate after multiple layers. When a single model is used for prediction, the prediction accuracy cannot be guaranteed, and at the moment, the combination of two models or multiple models can greatly improve the prediction accuracy. The hybrid model is widely applied to prediction of photovoltaic power generation. For example, in the document [5], the characteristics of extracting data space characteristics by a convolutional neural network and the characteristics of extracting time characteristics by a long-term and short-term memory network are combined, and photovoltaic power is predicted after an error inverse method and an optimized distributed gradient enhancement library machine learning model XBoost model are spliced in parallel. Due to the deviation of the historical data, the prediction accuracy of the model is reduced due to the excessive input amount, so that the data needs to be preprocessed. The conventional data processing mainly comprises normalization processing before inputting into the neural network model. But the simple data preprocessing cannot improve the prediction precision, and errors of the photovoltaic prediction model can be reduced by adding some data processing algorithms. At present, there are data processing methods such as Principal Component Analysis (PCA), pearson correlation coefficient, cluster analysis method, fuzzy recognition method, empirical Mode Decomposition (EMD) and decision tree theory. For example, document [6] proposes to establish an LSTM model of each modality after dividing the modality by a variable-mode decomposition VMD, and perform error compensation prediction on an error sequence, and obtain a more accurate prediction result by combining prediction powers of each modality and an error to sum up.
According to the method, whether the prediction model constructed by the photovoltaic output system meets the accuracy requirement along with the changes of the running time, the environment and the state needs to be judged through comparison with the physical entity, and when the accuracy requirement is not met, the parameters need to be continuously corrected through a self-learning self-updating mode, so that the synchronization with the physical entity is realized, namely the digital twin model. The model considers historical data and real-time data at the same time, has a self-learning function, and tracks the photovoltaic power generation physical entity, thereby realizing accurate power prediction. For example, in the document [8], after a model is built for photovoltaic power generation power prediction through a genetic algorithm and a back propagation neural network, real-time information of a photovoltaic cell and an environment is collected in real time, power prediction is performed, and finally, an actual power value under a similar working condition is obtained through similar meteorological search, and a preliminary prediction result is corrected. The digital twin technology is developed from the initial aerospace field to the current manufacturing industry, industry and other fields. The construction of the digital twin system mainly adopts 2 methods: firstly, the method is based on a mechanism model, but parameters in the mechanism model need to be continuously updated through learning; and secondly, establishing a non-mechanism model by using methods such as machine learning, data mining and the like, wherein in the process of establishing the model, a mechanism model is sometimes required to be combined or the mechanism model is taken as a basis.
[1] Jiangweining, et al, "photovoltaic power prediction method based on NWP irradiance slope conversion," Shandong university school newspaper (engineering edition) 51.05 (2021): 114-121.
[2] Time limin, et al, "short term prediction of photovoltaic power generation based on grey correlation analysis and GeoMAN model," electrotechnical bulletin 36.11 (2021): 2298-2305. Doi.
[3] Tanjin et al, "Adaboost weather clustering ultra-short-term prediction method for photovoltaic power generation of microgrid". The power system automation 41.21 (2017): 33-39.
[4] Weekly, et al, "short-term prediction of power generation in large photovoltaic power plants based on time series," Power technology 45.11 (2021): 1490-1494.
[5 Tang De Qing, zhuwu, and Houlin supra. "ultra-short term photovoltaic power prediction based on CNN-LSTM-XGboost model." Power technology 46.09 (2022): 1048-1052.
[6] Wan Fuzhong, wangshuafeng, and Zhangli, "photovoltaic power generation ultra-short term power prediction based on VMD-LSTM and error compensation," solar science 43.08 (2022): 96-103. Doi.
[8] Super short term prediction of photovoltaic power generation power [ J ] grid technology based on digital twin [ J ] grid technology, 2021,45 (04): 1258-1264. DOI.
Disclosure of Invention
The invention aims to provide a photovoltaic power generation power prediction method based on a combined machine learning model and application thereof, wherein photovoltaic prediction is realized through photovoltaic power generation power of empirical mode decomposition, long-term and short-term memory and support vector regression, and a superposition weight in the prediction model is corrected through actual output data of a photovoltaic system, so that a digital twin model for photovoltaic power generation power prediction is realized. The method comprises the steps of processing original photovoltaic output data through empirical mode decomposition and a sampling entropy method to obtain trend item data and fluctuation item data, predicting the trend item data through a long-term and short-term memory method, and predicting the trend item through a support vector regression method to obtain short-term photovoltaic output fluctuation item and trend item prediction results. And the weight required by superposition of the fluctuation item and the trend item is obtained by real-time optimization calculation of a particle swarm algorithm according to the actual power of the physical entity of the photovoltaic system.
After the trend item and the fluctuation item of the photovoltaic system output data are distinguished by combining an Empirical Mode Decomposition (EMD) algorithm and a Sampling Entropy (SE) algorithm, a long-short term memory (LSTM) method is applied to the construction of a trend item prediction model of photovoltaic power generation time sequence data, and a vector regression (SVR) algorithm is supported to construct a fluctuation item prediction learning model. In general, the output prediction result is obtained by overlapping the weight values of the trend item and the fluctuation item, the fluctuation influence of the digital twin physical entity is considered, and the actual photovoltaic output data of the physical entity is used for searching the optimal weight value to enable the prediction result to be more accurate. The method is analyzed and compared with the output of a prediction model only depending on the LSTM method and the output of an Empirical Mode Decomposition (EMD) + LSTM + SVR prediction model with superimposed equal weight values. The three photovoltaic system prediction results show that the prediction method for distinguishing the photovoltaic fluctuation item and the trend item after modal decomposition processing has better performance than the prediction method only relying on the long-term and short-term memory LSTM, and the digital twin method of the superposition weight corrected based on actual photovoltaic output data is more accurate than the prediction model prediction result of equal-weight superposition.
The invention provides a photovoltaic system output prediction method based on modal decomposition, a long-short term memory model and support vector regression, and provides a digital twin method for model self-updating, and a photovoltaic power generation power prediction digital twin model method based on a combined machine learning model, which comprises the following steps:
step 1: preprocessing photovoltaic output data to remove inconsistent, incomplete and noise-containing data;
step 2: analyzing the characteristics of photovoltaic output data, and extracting intrinsic mode vectors and residual terms of the power data of the photovoltaic system by adopting an Empirical Mode Decomposition (EMD) algorithm;
and step 3: calculating sampling entropy SE of each eigenmode vector of the photovoltaic output data;
and 4, step 4: distinguishing a trend item and a fluctuation item of the photovoltaic data, setting a threshold value by analyzing the sampling entropy of the whole photovoltaic output sequence, and superposing the eigenvector sampling entropies larger than the threshold value as the fluctuation item and superposing the eigenvector sampling entropies smaller than the threshold value as the trend item;
and 5: predicting the trend item by adopting a long-short term memory (LSTM) method;
step 6: predicting the fluctuation item by adopting a Support Vector Regression (SVR) method;
and 7: and constructing a digital twin model for photovoltaic power generation power prediction, and calculating and updating the predicted data superposition weight of the photovoltaic trend item and the fluctuation item in real time by using a particle swarm optimization PSO algorithm in combination with actual power data of a photovoltaic system.
Due to the adoption of the technical scheme, the invention has the following advantages:
1. the combined machine learning model is adopted to predict the photovoltaic output data, the efficiency is higher than that of a single prediction model, and the accuracy of the predicted data is improved.
2. And considering the accuracy of the power data of the photovoltaic system under different scales, predicting the volatility and the trend of the photovoltaic data by using corresponding methods respectively, so that the prediction result is more accurate.
And 3, by applying the concept of digital twinning, the constructed model is used for tracking the output condition of the physical entity and updating the weight coefficient, so that the photovoltaic power generation power prediction digital twinning model is more accurate.
Drawings
FIG. 1 is a diagram of modal decomposition results provided by the present invention;
FIG. 2 is a flow chart of sample entropy calculation provided by the present invention;
FIG. 3 is a power diagram of trend term and fluctuation term of a photovoltaic power plant of the present invention;
FIG. 4 is a plot of the long term short term memory LSTM neural network training process ios of the present invention;
FIG. 5 is a training flow chart of the weight updating particle swarm algorithm of the present invention;
FIG. 6 is a flow chart of photovoltaic power prediction provided by the present invention;
FIG. 7 is a graph of actual power of a photovoltaic power plant provided by the present invention;
FIG. 8 is a graph of the model fluctuation term prediction results provided by the present invention;
FIG. 9 is a graph of the model trend term prediction results provided by the present invention;
FIG. 10 is a graph of the prediction results of the hybrid model based on data preprocessing provided by the present invention;
FIG. 11 is a diagram illustrating the result of updating the weight values according to the present invention;
fig. 12 is a schematic structural diagram of an embodiment of an electronic device provided in the present invention.
Detailed Description
According to the method, an Empirical Mode Decomposition (EMD) algorithm and a Sampling Entropy (SE) algorithm are combined, after a trend item and a fluctuation item of output data of a photovoltaic system are distinguished, a long-term and short-term memory (LSTM) method is applied to the construction of a trend item prediction model of photovoltaic power generation time series data, and a vector regression (SVR) algorithm is supported to construct a fluctuation item prediction learning model. In general, the output prediction result is obtained by overlapping the weight values of the trend item and the fluctuation item, the fluctuation influence of the digital twin physical entity is considered, and the actual photovoltaic output data of the physical entity is used for searching the optimal overlapped weight value, so that the prediction result is more accurate. The method is analyzed and compared with the output of a prediction model only depending on the long-short term memory LSTM method and the output of an empirical mode decomposition + long-short term memory + support vector regression EMD + LSTM + SVR prediction model with equal weight value superposition. The three photovoltaic system prediction results show that the prediction method for distinguishing the photovoltaic fluctuation item and the trend item after modal decomposition processing has better performance than the prediction method only relying on the long-term and short-term memory LSTM, and the digital twin method of the superposition weight corrected based on actual photovoltaic output data is more accurate than the prediction model prediction result of equal-weight superposition.
The invention provides a photovoltaic system output prediction method based on modal decomposition, a long-short term memory model and support vector regression on the one hand, and provides a model self-updating digital twin method on the other hand, which comprises the following steps:
step 1: preprocessing the photovoltaic output data, and removing inconsistent, incomplete and noise-containing data:
checking whether the original photovoltaic data sequence has the following problems: (1) inconsistency problem, namely inconsistency condition occurs in data inclusion; (2) incomplete problem, missing attribute data of interest; (3) noise is a problem, and data having errors or abnormalities (deviation from an expected value) is present in the data.
For the problems, whether problem data exist is checked, and inconsistent data and data containing noise are eliminated. For incomplete data, linear interpolation of data is selected for supplementation. The linear interpolation method is to use a series of end-to-end linear connecting adjacent points in turn, and the height of the point in each line segment is used as the height value obtained by interpolation. With (x) i ,y i ) Indicating the previous end of a line segment, (x) i+1 ,y i+1 ) Indicating the latter end of the line segment, then for the line segment at (x) i ,x i+1 ) Points in the range with abscissa x have a height y defined as follows:
for example, historical data of photovoltaic power loss is obtained, and the relation between the data of temperature and illumination amplitude adjacent to unknown data and corresponding photovoltaic power data is matched through a linear interpolation method.
Step 2: analyzing the characteristics of photovoltaic output data, and extracting intrinsic mode vectors and residual terms of the power data of the photovoltaic system by using an Empirical Mode Decomposition (EMD) algorithm:
the EMD algorithm comprises the following specific steps:
1, taking the photovoltaic output data preprocessed in the step 1 as an input signal, and solving a maximum value signal and a minimum value signal of the input signal X (t);
2 structural Upper and lower envelope X u (t) and X l (t) calculating the sequence X a (t)：
X a ＝(X u (t)+X l (t))/2
3 judging step 2 to obtain X a (t) whether an eigenvector (IMF) condition is satisfied, if yes, IMF, otherwise, repeating 1 and 2;
4 calculating the signal R 1 And R1 is the value obtained by subtracting the IMF eigenvector obtained in the step 3 from the input signal:
R 1 ＝X(t)-IMF 1
replacing the original input signal X (t) with the residual signal R1, repeating the above steps 1-3 until the nth time the residual signal Rn obtained by subtracting the latest eigenmode vector IMF from the input signal is a monotone signal or only has an extreme point, and obtaining the following formula:
X(t)＝∑IMF i +R n
wherein X (t) is the original input signal, IMF i Rn is the residual signal for each eigenvector calculated.
The following two conditions are used in the step 3 of the EMD algorithm to determine whether the IMF is necessary and insufficient:
1 the number of the extreme points and zero-crossing points of the current signal should be equal or differ by one at most;
2 the mean value of the upper envelope line and the lower envelope line of the local maximum and the local minimum is 0;
note that: the above two requirements are necessary and insufficient conditions, that is, the IMF must satisfy the above two conditions, but the IMF does not have to satisfy the above two conditions.
And 3, step 3: and calculating the sampling entropy SE of each eigenmode vector of the photovoltaic output data.
And 3, calculating the sampling entropy of each eigenmode vector according to the step 3:
sample entropy measures the complexity of a time series by measuring the magnitude of the probability of generating a new pattern in a signal, the greater the probability of generating a new pattern, the greater the complexity of the series. The sample entropy does not depend on the data length, and the sample entropy has better consistency. The lower the value of the sample entropy, the higher the sequence self-similarity, the larger the value of the sample entropy, the more complex the sample sequence. The calculation method comprises the following steps:
let us assume that a time series { x (N) } consisting of N data is represented as { x (N) } = x (1), x (2), \8230; x (N):
1 forming a set of vector sequences of dimension m from sequence numbers, X m (1),…X m (N-m + 1) wherein X m (i) Is a sequence of m-dimensional vectors, X m (i) = { x (i), x (i + 1) \8230: (i + m-1) },1 ≦ i ≦ N-m +1. These vectors represent m consecutive values of x starting from point i;
2 definition of vector X m (i) And X m (j) Distance d [ X ] between m (i),X m (j)]As the absolute value of the maximum difference between the two corresponding elements, i.e.
d[X m (i),X m (j)]＝max k＝0,K,m-1 (|x(i+k)-x(j+k)|)
3 for a given vector X m (i) Statistics of X m (i) And X m (j) The number of j (j is more than or equal to 1 and less than or equal to N-m, j is not equal to i) with the distance between the two is less than or equal to r and is marked as B i . For 1. Ltoreq. I.ltoreq.N-m, defined:
wherein B is m i (r) is X with a distance not greater than distance r m (i) And X m (j) The number of (j is more than or equal to 1 and less than or equal to N-m, j is not equal to i).
4 definition of B (m) (r) m-dimension X with distance not greater than distance r m (i) And X m (j) The average number of dimensions (1. Ltoreq. J. Ltoreq.N-m, j. Noteq. I) of the number of (1 ≦ j. Ltoreq. N-m, j. Noteq. I) is the probability that two sequences match m points with a similarity tolerance r, given by:
5 increasing dimension to m +1, calculating X m+1 (i) And X m+1 (j) (j is more than or equal to 1 and less than or equal to N-m, j is not equal to i) the number of distances less than or equal to r is marked as A i 。A i m (r) is defined as:
wherein A is m i (r) X in m +1 dimensions at a distance of not more than r m+1 (i) And X m+1 (j) (j is not less than 1 and not more than N-m, j is not equal to i))。
6 definition A m (r) is m + 1-dimensional X with a distance not greater than distance r m (i) And X m (j) The average number of dimensions (j is more than or equal to 1 and less than or equal to N-m, j is not equal to i), i.e. the probability that two sequences match m +1 points, is:
thus, B m (r) is the probability that two sequences match m points with a similarity tolerance of r, and A m (r) is the probability that two sequences match m +1 points, and the sample entropy is defined as:
when N is finite, it is estimated by the following equation:
and 4, step 4: the method comprises the following steps of distinguishing a trend item and a fluctuation item of photovoltaic data, setting a threshold value by analyzing the sampling entropy of the whole photovoltaic output sequence, superposing the eigenvector sampling entropies larger than the threshold value as the fluctuation item, and superposing the eigenvector sampling entropies smaller than the threshold value as the trend item:
and selecting each eigenmode vector as a fluctuation item and a trend item according to the step 4:
and calculating the output data of the photovoltaic system and the sampling entropy of each eigenvector, and setting a proper threshold u by an observation method. If the sampling entropy value of the eigenvector is more than or equal to u, the eigenvector is taken as a fluctuation item element; and if the sampling entropy value of the eigenvector is smaller than u, the eigenvector is taken as a trend item element.
The method is characterized in that M eigenvectors are decomposed by assuming the output data mode of the photovoltaic system, wherein M 1 If the sum is compared with a set threshold value u through SE calculation, and if the sum is less than the threshold value u, the sum is a fluctuation item element, (M-M) 1 ) And each is a trend item element. The fluctuation term and the trend term are calculated as the superposition of the elements,as shown in the following formula:
combinations of eigenvectors smaller than a threshold u as fluctuation terms, i.e. X flu With the remaining eigenvectors as trend terms, i.e., X ten 。
And 5: predicting the trend item by using a photovoltaic system generated power trend item long-term and short-term memory LSTM prediction method:
and (3) predicting long-term and short-term memory of the fluctuation item:
performing primary prediction on photovoltaic output power by adopting a prediction program, and completing the prediction in two stages of a training process and a prediction process; the trend item data is divided into training subdata and testing subdata, the data is standardized to construct a predicted learning model, and parameter values of the learning model are selected through a training data minimization loss function by adopting an Adam optimizer.
The photovoltaic system trend term long-term and short-term memory LSTM prediction method comprises two stages:
1 first stage, training Process
And 4, preprocessing the trend item and fluctuation item data set obtained in the step 4, intercepting the data into a set of n supervised learning formats, zooming the data into a [0,1] interval, dividing 2/3 data in the data to be used as training, setting a prediction step length to be n, namely predicting one data through n data, and substituting the data into the training set to construct the long-short term memory LSTM network.
2 second stage, prediction Process
And based on the long-short term memory LSTM network constructed in the last stage, substituting the rest 1/3 of data into the constructed long-short term memory LSTM network for prediction. And merging the prediction results of the fluctuation item and the trend item.
The LSTM can extract the hidden characteristics of time series data, and the algorithm comprises three gates: forgetting gate, input gate and output gate. Wherein the output formula of the forgetting gate is as follows:
f (t )＝σ(W i [h (t-1) ,x (t) ]+b i )
where σ () represents a sigmoid function, f (t) Representing the forgetting probability of the previously hidden cell state. W f A weight representing a forgetting gate, and b f Indicating a forgotten door bias.
The output formula of the input gate is expressed as:
i (t) ＝σ(W i [h (t-1) ,x ( t ) ]+b i )
a (t) ＝tanh(W a [h (t-1) ,x (t) ]+b a )
wherein, W i Representing the weight of the input gate, and b i Representing the bias of the input gate, tanh () representing the activation function, W a A weight representing the state of cell renewal, and b a Bias representing the state of cell renewal i (t) For input of gate output, a (t) Is a signal of whether the cell is renewing. The updated formula for the cell state is:
C (t) ＝C (t-1) e f (t) +i (t) e a (t)
wherein, C (t) The state of the cell at time t, C (t-1) Is the state of the cell at time t; c (t-1) ⊙f (t) Deciding whether to save the original state of the t-1 time state, i (t) ⊙a (t) Cell status indicating whether to update the t-time status.
The output formula is expressed as:
a (t) ＝σ(W o [h (t-1) ,x (t) ]+b o )
h (t) ＝o (t) e tanh(C (t) )
wherein, a (t) To output the formula, W o Representing the weight of the output gate, and b o Denotes the offset of the output gate, h (t) Indicating the delivery status at time t.
After format processing before training is carried out on the photovoltaic power historical prediction data, a training set and a test set are divided, the training set data is used for assisting in building the LSTM network, and the test set data is substituted into the long-term and short-term memory LSTM network for prediction to obtain the prediction value of the test set.
Step 6: method for predicting fluctuation item by using Support Vector Regression (SVR)
Carrying out support vector regression prediction on photovoltaic output fluctuation data, wherein the step is completed in two stages of a training process and a prediction process; the fluctuation item data is divided into training subdata and testing subdata, the data is standardized to construct a predicted learning model, a radial basis function rbf is adopted as a hidden layer neuron activation function, and parameter values of the learning model are selected through a training data minimum loss function.
The photovoltaic system fluctuation term support vector regression SVR prediction method comprises two stages.
1 first stage, training process.
And 6, preprocessing the fluctuation item data set obtained in the step 6, and intercepting the data into a specified supervised learning format. Substituting all training set data to construct a Support Vector Regression (SVR) regression model.
2 second stage, prediction Process
And substituting the prediction time into the constructed model to perform regression fitting prediction based on the Support Vector Regression (SVR) regression model constructed in the last stage.
The SVR method requires defining a constant ε>0, for a certain point (x) i ,y i ) If the following equation is satisfied:
|y i -w·φ(x i )-b|≤ε
no loss is indicated, and if the above formula is not satisfied, a loss is indicated. Where w and b represent weights and offsets. The SVR model calculates the loss only above a threshold epsilon and optimizes the model by maximizing the width of the spacer band and minimizing the loss. Because the fluctuation item is predicted according to a fitting rule in a certain range, after the training set and the test set are divided, the SVR is subjected to fitting regression by using historical data, and the predicted value of the fluctuation item is obtained by predicting by using the data of the test set.
And 7: and constructing a digital twin model for photovoltaic power generation power prediction. And (4) calculating and updating the superposition weight of the predicted data of the photovoltaic trend item and the fluctuation item in real time by using a PSO algorithm in combination with the actual power data of the photovoltaic system.
And calculating the superposition weight of the updated trend item and the fluctuation item by a PSO algorithm in combination with the actual output of the physical entity of the photovoltaic system, so as to realize the construction of the photovoltaic power generation power prediction digital twin model.
The core formula of the PSO algorithm is as follows:
V id (t+1)＝wV id +c 1 *rand 1 ()(Pbest id (t)-X id (t))+c 2 *rand 2 ()(Gbest id (t)-X id (t))
(21)
X id (t+1)＝X id (t)+V id (t+1)
wherein w, c 1 And c 2 Is a set value, w is an inertial weight, c 1 And c 2 Is a learning factor; pbest id (t) and Gbest id (t) two position parameters at time t are updated, V id To update the speed parameter, V id (t) is the velocity at time t, X id To update the location parameters, X id (t) is the position at time t. rand 1 () And rand 2 () Is a random probability. In the experiment, the optimal weight values of the two methods are trained by using the data of the test set and the data error minimization of the weight value optimization as indexes according to the prediction results obtained by the long-term and short-term memory LSTM and the support vector regression SVR method. The quality of the prediction model is quantified by using a statistical index, and the performance of the prediction model is evaluated by using a Root Mean Square Error (RMSE), wherein a calculation formula is as follows:
in the formula y t Is the actual value of the one or more parameters,
are corresponding estimates and n is the number of measurements.
The photovoltaic output digital twin model system based on modal decomposition, long-term and short-term memory, support vector regression and particle swarm optimization is used for realizing the prediction method. The device comprises a model building module, a power prediction module and a digital twin realization module.
The model building module is used for data preprocessing and building two model networks for predicting a trend item and a fluctuation item of the photovoltaic data. The power prediction module is used for predicting photovoltaic power based on the prediction network. The digital twin realization module is used for constructing a photovoltaic power generation power prediction digital twin model.
Examples
The method takes the photovoltaic output data of a certain Changzhou photovoltaic power station 2018 in 2 months as an experimental object to show an embodiment.
The method comprises the steps of carrying out data elimination on inconsistent conditions and noise-containing conditions of photovoltaic power data, and carrying out linear data interpolation on data under incomplete conditions. The final photovoltaic sequence is shown in figure 7.
And performing empirical mode EMD on the two photovoltaic integral sequences to obtain 8 eigenmode vectors and a residual value, as shown in figure 1.
And step three, carrying out decomposition processing on the eigenvectors obtained in the step two, wherein a calculation flow chart of the sampling entropy is shown as the attached figure 2, and the output data of the photovoltaic system and the results of the eigenvectors after modal decomposition are shown as the following table:
and step four, setting a threshold value u by an observation method, namely, the u is 0.12. And taking the superposition item of the IMF2 and the IM3 as a trend item, and taking other vectors as fluctuation items. Wherein figure 3 shows the variation of the fluctuation term and the trend term.
And step five, predicting the trend item of the superposition of the eigenvectors IMF2 and IMF3 by using a long-short term memory LSTM method, wherein the attached figure 4 is a training process diagram of the long-short term memory LSTM prediction. FIG. 8 is a long short term memory LSTM test set prediction graph.
And step six, predicting the fluctuation item constructed by the residual eigenvector by using a Support Vector Regression (SVR) method, wherein the attached figure 9 is a prediction curve of the SVR for the fluctuation item. The upper part of fig. 9 is the fitting case of 100 to 150 points in the test set, and the lower part is the whole case in the test set.
And seventhly, realizing the digital twin model by using a Particle Swarm Optimization (PSO). FIG. 11 shows the optimal weight distribution of the LSTM, and the weight of the LSTM subtracted by 1 is the weight of the SVR. The final predicted effect is shown in figure 10.
The platform of the project is shown in fig. 12 and is composed of a memory and a processor, and the overall experimental block diagram is shown in fig. 6.
The prediction model is compared with a model which is only predicted by using a long-short term memory (LSTM) model and is constructed by using equal weight values for experiments, and the experimental result is shown in the following table:
from the above table, it can be seen that the hybrid model based on data preprocessing is more accurate than the prediction result relying only on the long-short term memory LSTM model. And the mixed model after weight updating has better training effect than the mixed model with equivalent superposition.
The defect that the accuracy of a single prediction model is not high in the background technology is finally overcome through the steps 1-7, and the effect that the photovoltaic power prediction model improves the prediction accuracy is achieved.
The invention aims to provide a photovoltaic power generation power prediction digital twin model based on a combined machine learning model, wherein photovoltaic prediction is realized through photovoltaic power generation power of empirical mode decomposition, long-term and short-term memory and support vector regression, and a superposition weight in the prediction model is corrected through actual output data of a photovoltaic system, so that the photovoltaic power generation power prediction digital twin model is realized. The method comprises the steps of processing original photovoltaic output data through empirical mode decomposition and a sampling entropy method to obtain trend item data and fluctuation item data, predicting the trend item data through a long-term and short-term memory method, and predicting the trend item through a support vector regression method to obtain short-term photovoltaic output fluctuation item and trend item prediction results. And the weight value required by superposition of the fluctuation item and the trend item is obtained by real-time optimization calculation of a particle swarm algorithm according to the actual power of the physical entity of the photovoltaic system. The invention belongs to the technical field of intelligent electric power.
The foregoing shows and describes the general principles, principal features, and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are merely illustrative of the principles of the invention, but that various changes and modifications may be made without departing from the spirit and scope of the invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (9)
1. A photovoltaic power generation power prediction method based on a combined machine learning model is characterized by comprising the following steps: the method comprises the following steps:
step 1: preprocessing photovoltaic output data, and removing inconsistent, incomplete and noise-containing data;
step 2: analyzing the characteristics of photovoltaic output data, and extracting intrinsic mode vectors and residual terms of the power data of the photovoltaic system by adopting an Empirical Mode Decomposition (EMD) algorithm;
and 3, step 3: calculating sampling entropy SE of each eigenmode vector of the photovoltaic output data;
and 4, step 4: distinguishing a trend item and a fluctuation item of the photovoltaic data, setting a threshold value by analyzing the sampling entropy of the whole photovoltaic output sequence, and superposing the eigenvector sampling entropies larger than the threshold value as the fluctuation item and superposing the eigenvector sampling entropies smaller than the threshold value as the trend item;
and 5: predicting the trend item by adopting a long-short term memory (LSTM) method;
step 6: predicting the fluctuation item by adopting a Support Vector Regression (SVR) method;
and 7: and constructing a digital twin model for photovoltaic power generation power prediction, and calculating and updating the predicted data superposition weight of the photovoltaic trend item and the fluctuation item in real time by using a particle swarm optimization PSO algorithm in combination with actual power data of a photovoltaic system.
2. The combined machine learning model-based photovoltaic power generation power prediction method according to claim 1, characterized by: the step 1 further comprises the following steps: the original photovoltaic data sequence was checked for the following problems: (1) inconsistency problem, namely inconsistency condition occurs in data inclusion; (2) incomplete problem, missing attribute data of interest; (3) noise is a problem and there are errors, or anomalies, in the data, i.e., data that deviate from the expected value.
3. The combined machine learning model-based photovoltaic power generation power prediction method of claim 1, characterized by: the EMD algorithm comprises the following specific steps:
1) Taking the photovoltaic output data preprocessed in the step 1 as input signals, and solving maximum and minimum signals of the input signals X (t);
2) Constructing the upper and lower envelope X u (t) and X l (t) calculating the sequence X a (t)：
X a ＝(X u (t)+X l (t))/2
3) Judging the sequence X calculated in the second step a (t) whether an eigenvector (IMF) condition is satisfied, if yes, IMF, otherwise, repeating 1 and 2;
4) Calculating a residual signal R 1 ，R 1 Subtracting the value of the IMF eigenvector obtained in the third step from the input signal:
R 1 ＝X(t)-IMF 1
from the residual signal R 1 Replacing the original input signal X (t), repeating the steps 1-3 until the nth time the residual signal R obtained by subtracting the latest obtained intrinsic mode vector IMF from the input signal n For monotonous signals or presence of only oneUp to the extreme, the following formula is obtained:
X(t)＝∑IMF i +R n 。
wherein X (t) is the original input signal, IMF i For each eigenvector calculated, R n Is a residual signal.
4. The combined machine learning model-based photovoltaic power generation power prediction method of claim 1, characterized by: the long-short term memory LSTM algorithm comprises the following specific steps:
assume that a time series { x (N) } = x (1), x (2), \ 8230; x (N):
1) Forming a series of vectors of dimension m, i.e. X, from the sequence numbers m (1),…X m (N-m + 1) wherein the ith vector sequence is X m (i) = { x (i), x (i + 1) \ 8230, x (i + m-1) },1 ≦ i ≦ N-m +1. These vectors represent m consecutive values of x starting from the ith point;
2) Definition vector X m (i) And X m (j) Distance d [ X ] between m (i),X m (j)]Is the absolute value of the maximum difference between the two corresponding elements, as shown in the following equation:
d[X m (i,X m (j)]＝max k＝0,K,m-1 (|x(i+k)-x(j+k)|)
3) For a given X m (i) Statistics of X m (i) And X m (j) The number of j (j is more than or equal to 1 and less than or equal to N-m, j is not equal to i) with the distance between the two is less than or equal to the distance r, and is marked as B i . For 1. Ltoreq. I.ltoreq.N-m, defined:
wherein B is m i (r) is X with a distance not greater than distance r m (i) And X m (j) The number of (j is more than or equal to 1 and less than or equal to N-m, j is not equal to i);
4) Definition B (m) (r) m-dimension X with distance not greater than distance r m (i) And X m (j) The average number of dimensions (1. Ltoreq. J. Ltoreq.N-m, j. Noteq. I) of the number of (1 ≦ j. Ltoreq. N-m, j. Noteq. I) is the probability that two sequences match m points with a similarity tolerance r, given by:
5) Add dimension to m +1, calculate X m+1 (i) And X m+1 (j) (j is more than or equal to 1 and less than or equal to N-m, j is not equal to i) the number of the distances less than or equal to r is marked as A i 。A i m (r) is defined as:
wherein A is m i (r) X in m +1 dimensions at a distance of not more than r m+1 (i) And X m+1 (j) The number of (j is more than or equal to 1 and less than or equal to N-m, j is not equal to i);
6) Definition A m (r) is m + 1-dimensional X with a distance not greater than the distance r m (i) And X m (j) The average number of dimensions (j is greater than or equal to 1 and less than or equal to N-m, j is not equal to i), i.e. the probability that two sequences match m +1 points, is:
thus, B m (r) is the probability that two sequences match m points with a similarity tolerance r, and A m (r) is the probability that two sequences match m +1 points, and the sample entropy is defined as:
when N is finite, it is estimated by the following equation:
5. the combined machine learning model-based photovoltaic generated power prediction digital twin model method of claim 1, characterized by: the step 4 further comprises the following steps:
the method is characterized in that M eigenvectors are decomposed by assuming the output data mode of the photovoltaic system, wherein M 1 The element (M-M) is a fluctuation item element when the sum is compared with a set threshold u through the calculation of sampling entropy SE and is smaller than the threshold 1 ) Each is a trend item element; the fluctuation term and the trend term are calculated as the superposition of the elements as shown in the following formula:
combinations of eigenvectors smaller than a threshold u as fluctuation terms, i.e. X flu With the remaining eigenvectors as trend terms, i.e., X ten 。
6. The combined machine learning model-based photovoltaic generated power prediction digital twin model method of claim 5, characterized by: the photovoltaic system trend item generated power long-term and short-term memory LSTM prediction method comprises two stages:
first stage, training process
Preprocessing the trend item and fluctuation item data set obtained in the step 4, intercepting the data into a set of n supervised learning formats, zooming the data into a [0,1] interval, dividing 2/3 data in the data as training, setting a prediction step length as n, namely predicting one data through n data, and substituting the data into the training set to construct a long-short term memory (LSTM) network;
second stage, prediction process
And based on the long-short term memory LSTM network constructed in the last stage, substituting the rest 1/3 of data into the constructed long-short term memory LSTM network for prediction. Combining the prediction results of the fluctuation item and the trend item;
the long-short term memory LSTM can extract the hidden characteristics of time series data, and the algorithm comprises three gates: forgetting gate, input gate and output gate. Wherein the output formula of the forgetting gate is as follows:
f (t) ＝σ(W i [h (t-1) ,x (t) ]+b i )
where σ () represents a sigmoid function, f (t) Representing the forgetting probability of the previously hidden cell state. W f A weight representing a forgetting gate, and b f A bias indicating a forgetting gate;
the output formula of the input gate is expressed as:
i (t) ＝σ(W i [h (t-1) ,x (t) ]+b i )
a (t) ＝tanh(W a [h (t-1) ,x (t) ]+b a )
wherein, W i Representing the weight of the input gate, and b i Representing the bias of the input gate, tanh () representing the activation function, W a A weight representing the state of cell renewal, and b a Bias representing the state of cell renewal i (t) For input gate output, a (t) A signal whether the cell is renewing;
the updated formula for the cell state is:
C (t) ＝C (t-1) ef (t) +i (t) ea (t)
wherein, C (t) The state of the cell at time t, C (t-1) Is the state of the cell at time t; c (t-1) ⊙f (t) Deciding whether to save the original state of the t-1 time state, i (t) ⊙a (t) A cell state indicating whether to update the t-time state;
the output formula is expressed as:
a (t) ＝σ(W o [h (t-1 ) , x (t) ]+b o )
h (t) ＝o (t) e tanh(C (t) )
wherein, a (t) To output the formula, W o Representing the weight of the output gate, and b o Denotes the offset of the output gate, h t Indicating the delivery status at time t.
7. The combined machine learning model-based photovoltaic generated power prediction digital twin model method according to claim 1, characterized by: calculating the superposition weight of the updating trend item and the fluctuation item through a PSO algorithm to realize the construction of a photovoltaic power generation power prediction digital twin model;
the core formula of the PSO algorithm is as follows:
V id (t+1)＝wV id +c 1 *rand 1 ( )(Pbest id (t)-X id (t))+c 2 *rand 2 ( )(Gbest id (t)-X id (t))
(17)
X id (t+1)＝X id (t)+V id (t+1) (18)
wherein w, c 1 And c 2 Is a set value, w is an inertial weight, c 1 And c 2 Is a learning factor; pbest id (t) and Gbest id (t) two position parameters at time t are updated, V id To update the speed parameter, V id (t) is the velocity at time t, X id To update the location parameters, X id (t) is the position at time t. rand 1 () And rand 2 () Is a random probability; and training the optimal weight values of the two methods by using the data of the test set and the data error minimization of the weight value optimization as an index according to the prediction results obtained by the long-short term memory LSTM and the support vector regression SVR method.
8. A non-volatile storage medium, comprising a stored program, wherein the program, when executed, controls an apparatus in which the non-volatile storage medium is located to perform the method of any one of claims 1 to 7.
9. An electronic device comprising a processor and a memory; the memory has stored therein computer readable instructions for execution by the processor, wherein the computer readable instructions when executed perform the method of any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211325112.5A CN115860177A (en) | 2022-10-27 | 2022-10-27 | Photovoltaic power generation power prediction method based on combined machine learning model and application thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211325112.5A CN115860177A (en) | 2022-10-27 | 2022-10-27 | Photovoltaic power generation power prediction method based on combined machine learning model and application thereof |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115860177A true CN115860177A (en) | 2023-03-28 |
Family
ID=85661904
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211325112.5A Pending CN115860177A (en) | 2022-10-27 | 2022-10-27 | Photovoltaic power generation power prediction method based on combined machine learning model and application thereof |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115860177A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116187205A (en) * | 2023-04-24 | 2023-05-30 | 北京智芯微电子科技有限公司 | Running state prediction method and device for digital twin body of power distribution network and training method |
| CN117117923A (en) * | 2023-10-19 | 2023-11-24 | 深圳市百酷新能源有限公司 | Big data-based energy storage control grid-connected management method and system |
| CN117713221A (en) * | 2024-02-01 | 2024-03-15 | 深圳戴普森新能源技术有限公司 | Micro-inversion photovoltaic grid-connected optimization system |
-
2022
- 2022-10-27 CN CN202211325112.5A patent/CN115860177A/en active Pending
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116187205A (en) * | 2023-04-24 | 2023-05-30 | 北京智芯微电子科技有限公司 | Running state prediction method and device for digital twin body of power distribution network and training method |
| CN116187205B (en) * | 2023-04-24 | 2023-08-15 | 北京智芯微电子科技有限公司 | Running state prediction method and device for digital twin body of power distribution network and training method |
| CN117117923A (en) * | 2023-10-19 | 2023-11-24 | 深圳市百酷新能源有限公司 | Big data-based energy storage control grid-connected management method and system |
| CN117117923B (en) * | 2023-10-19 | 2024-04-05 | 深圳市百酷新能源有限公司 | Big data-based energy storage control grid-connected management method and system |
| CN117713221A (en) * | 2024-02-01 | 2024-03-15 | 深圳戴普森新能源技术有限公司 | Micro-inversion photovoltaic grid-connected optimization system |
| CN117713221B (en) * | 2024-02-01 | 2024-04-16 | 深圳戴普森新能源技术有限公司 | Micro-inversion photovoltaic grid-connected optimization system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113962364B (en) | Multi-factor power load prediction method based on deep learning | |
| CN109711620B (en) | Short-term power load prediction method based on GRU neural network and transfer learning | |
| CN111260136A (en) | Building short-term load prediction method based on ARIMA-LSTM combined model | |
| CN115860177A (en) | Photovoltaic power generation power prediction method based on combined machine learning model and application thereof | |
| CN111027772A (en) | Multi-factor short-term load prediction method based on PCA-DBILSTM | |
| CN113537582B (en) | Photovoltaic power ultra-short-term prediction method based on short-wave radiation correction | |
| Liu et al. | Heating load forecasting for combined heat and power plants via strand-based LSTM | |
| CN116167527B (en) | Pure data-driven power system static safety operation risk online assessment method | |
| CN114218872B (en) | DBN-LSTM semi-supervised joint model-based residual service life prediction method | |
| CN112329990A (en) | User power load prediction method based on LSTM-BP neural network | |
| CN113554466A (en) | Short-term power consumption prediction model construction method, prediction method and device | |
| CN113011680A (en) | Power load prediction method and system | |
| CN115115125A (en) | Photovoltaic power interval probability prediction method based on deep learning fusion model | |
| CN115796004A (en) | Photovoltaic power station ultra-short term power intelligent prediction method based on SLSTM and MLSTNet models | |
| CN116341613A (en) | Ultra-short-term photovoltaic power prediction method based on Informar encoder and LSTM | |
| Alharbi et al. | Short-term wind speed and temperature forecasting model based on gated recurrent unit neural networks | |
| CN113610328A (en) | Power generation load prediction method | |
| Manoj et al. | FWS-DL: forecasting wind speed based on deep learning algorithms | |
| Zhang et al. | Inferential statistics and machine learning models for short-term wind power forecasting | |
| Yuan et al. | Short-term wind speed forecasting using STLSSVM hybrid model | |
| CN115481788A (en) | Load prediction method and system for phase change energy storage system | |
| Ding et al. | A statistical upscaling approach of region wind power forecasting based on combination model | |
| CN114819395A (en) | Industry medium and long term load prediction method based on long and short term memory neural network and support vector regression combination model | |
| Tang et al. | Traffic flow anomaly detection based on robust ridge regression with particle swarm optimization algorithm | |
| CN113191526A (en) | Short-term wind speed interval multi-objective optimization prediction method and system based on random sensitivity |
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 |