CN105809264B - Power load prediction method and device - Google Patents
Power load prediction method and device Download PDFInfo
- Publication number
- CN105809264B CN105809264B CN201410838273.3A CN201410838273A CN105809264B CN 105809264 B CN105809264 B CN 105809264B CN 201410838273 A CN201410838273 A CN 201410838273A CN 105809264 B CN105809264 B CN 105809264B
- Authority
- CN
- China
- Prior art keywords
- historical data
- electrical load
- data
- predicting
- belief network
- 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.)
- Active
Links
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The invention provides a power load prediction method and a power load prediction device. The method comprises the following steps: training a first deep belief network model for predicting the electrical load from the historical data of the electrical load with the historical data of the electrical load, thereby producing a first prediction result for predicting a future electrical load; training a second deep belief network model that predicts the electrical load from the historical data of the electrical load and the historical data of the predefined dominant impact factor with the historical data of the electrical load and the historical data of the predefined dominant impact factor, thereby producing a second prediction result for predicting a future electrical load; and predicting the future electric load according to the first prediction result and the second prediction result. The invention improves the accuracy of power load prediction, especially under the condition of partial missing of historical data or huge historical data.
Description
Technical Field
The invention relates to the field of intelligent prediction, in particular to a power load prediction method and a power load prediction device.
Background
Power forecasting is an important task in the operation and planning of electric utilities today. Only with a reasonable prediction of future electrical loads can a reasonable power planning be performed. However, in the case where the history information is limited or the amount of history information data is large, it is difficult to accurately predict the electric load.
Disclosure of Invention
In view of the above, one of the problems solved by the embodiments of the present invention is to improve the accuracy of the power load prediction.
According to an embodiment of the present invention, there is provided an electric load prediction method including: training a first deep belief network model for predicting the electrical load from historical data of the electrical load with the historical data of the electrical load, thereby producing a first prediction result for predicting a future electrical load; training a second deep belief network model that predicts the electrical load from the historical data of the electrical load and the historical data of the predefined dominant impact factor with the historical data of the electrical load and the historical data of the predefined dominant impact factor, thereby producing a second prediction result for predicting a future electrical load; and predicting the future electric load according to the first prediction result and the second prediction result.
Optionally, the step of training the first deep belief network model for predicting the electrical load from the historical data of the electrical load with the historical data of the electrical load further comprises: in the case where the historical data of the electric load is partially missing, the missing part is predicted from the electric load data before the missing part by the first deep belief network model and the historical data of the electric load is complemented.
Optionally, the method further comprises: and training a third deep belief network model for predicting the missing part according to the predefined dominant influence factor before the missing part and completing the historical data of the predefined dominant influence factor in the case that the historical data of the predefined dominant influence factor is missing part by using the historical data of the predefined dominant influence factor, thereby completing the historical data of the predefined dominant influence factor in the case that the historical data of the predefined dominant influence factor is partially missing.
Optionally, each of the first, second, and third deep belief network models includes a plurality of layers of logical nodes, wherein a lowest layer of logical nodes of the plurality of layers of logical nodes receives the input data, logical nodes of other layers except the lowest layer of logical nodes are related to logical nodes of a next layer, and the relationship is expressed by a coefficient, so that each of the first, second, and third deep belief network models is expressed as a coefficient matrix, and the training process of each of the first, second, and third deep belief network models includes: obtaining coefficients in a coefficient matrix from the input data for the multilayer logic nodes from bottom to top; coefficients in the coefficient matrix are verified from top to bottom for the multi-level logical nodes by backward fitting.
Optionally, the coefficient values of the first, second and third depth belief network models are obtained by greedy unsupervised feature learning.
Optionally, the historical data of the electrical load is normalized historical data of the electrical load, and the historical data of the predefined dominant impact factor is normalized historical data of the predefined dominant impact factor.
Optionally, the predefined dominant impact factor comprises at least one of weather data, temperature data, human context data.
An embodiment of the present invention provides an electrical load prediction apparatus including: a first training unit configured to train a first deep belief network model for predicting the electrical load from the historical data of the electrical load with the historical data of the electrical load, thereby generating a first prediction result for predicting a future electrical load; a second training unit configured to train a second deep belief network model predicting the electrical load from the historical data of the electrical load and the historical data of the predefined dominant impact factor with the historical data of the electrical load and the historical data of the predefined dominant impact factor, thereby generating a second prediction result for predicting the future electrical load; a prediction unit configured to predict a future electrical load based on the first prediction result and the second prediction result.
Optionally, the first training unit is further configured to predict the missing part from power load data before the missing part with the first deep belief network model and to complement the historical data of the power load in case the historical data of the power load is partially missing.
Optionally, the electrical load prediction apparatus further comprises: a third training unit configured to train a third deep belief network model for predicting the missing part from the predefined dominant impact factor before the missing part and complementing the historical data of the predefined dominant impact factor in case the historical data of the predefined dominant impact factor is partially missing, with the historical data of the predefined dominant impact factor, thereby complementing the historical data of the predefined dominant impact factor in case the historical data of the predefined dominant impact factor is partially missing.
Optionally, each of the first, second, and third deep belief network models includes a plurality of layers of logical nodes, wherein a lowest layer of the plurality of layers of logical nodes receives the input data, and logical nodes of other layers except the lowest layer of logical nodes are related to logical nodes of a next layer, and the relationship is represented by a coefficient, so that each of the first, second, and third deep belief network models is represented as a coefficient matrix, and the training process of each of the first, second, and third deep belief network models includes: obtaining coefficients in a coefficient matrix from the input data for the multilayer logic nodes from bottom to top; coefficients in the coefficient matrix are verified from top to bottom for the multi-level logical nodes by backward fitting.
Optionally, the coefficient values of the first, second and third depth belief network models are obtained by greedy unsupervised feature learning.
Optionally, the historical data of the electrical load is normalized historical data of the electrical load, and the historical data of the predefined dominant impact factor is normalized historical data of the predefined dominant impact factor.
Optionally, the predefined dominant impact factor comprises at least one of weather data, temperature data, human context data.
According to the embodiment of the invention, the first deep belief network model for predicting the power load according to the historical data of the power load is adopted, and the second deep belief network model for predicting the power load according to the historical data of the power load and the historical data of the predefined dominant influence factor is also adopted, so that the prediction result not only reflects the change trend of the power load, but also reflects the change trend of the influence of the relevant dominant influence factors (such as environment and weather) on the power load, and the prediction result is more objective, and the prediction precision is improved.
Since the first deep belief network model is a model for predicting the power load from the historical data of the power load in one embodiment of the present invention, it reflects the trend of the power load itself to change. The tendency of the electrical load to change itself is often the result of the synergy of all the influencing factors influencing the electrical load. Many of these influencing factors are incidental factors. Therefore, the model including a great number of accidental factors is used for prediction, so that the prediction result is always accidental. Some of the more important influencing factors are often buried in a large number of accidental influencing factors, making the prediction inaccurate. The second deep belief network model well makes up the defect, focuses on investigating the influence of some predefined dominant influence factors on the prediction result, eliminates the influence of contingency, and enables some factors which possibly play a decisive role in the prediction result to be reflected in the prediction result. By combining the two models, the influence of the contingency in the prediction result can be reduced, the influence of certain contingency in the absence of the prediction result can be objectively acknowledged, and the prediction precision is improved.
In addition, because the deep belief network model is adopted in the embodiment of the invention, the deep belief network model is a multilayer nonlinear network structure, a simple function expression and a complex relation expressed by some parameters can be used, and the processing precision and efficiency when the method faces huge historical data can be particularly improved. Therefore, the accuracy and the efficiency of predicting the power load are improved by adopting the deep belief network model.
In addition, the coefficient value of the deep belief network model adopted in the embodiment of the invention is obtained by greedy unsupervised feature learning, and the greedy unsupervised feature learning has the characteristic that the characteristic and the complex pattern hidden behind the data can be clearly expressed, so that even if part of historical data is lost, the data-backed relation can still be found by finding the data-backed relation from the historical data of incomplete power load, and accurate prediction is realized.
Drawings
Other features, advantages and benefits of the present invention will become more apparent from the following detailed description of the invention when taken in conjunction with the accompanying drawings.
Fig. 1 shows a flow chart of an electrical load prediction method according to an embodiment of the invention.
Fig. 2 shows a flow chart of an electrical load prediction method according to another embodiment of the invention.
Fig. 3 shows the structure of a standard deep belief network model.
FIG. 4 illustrates an example of first and second deep belief network models employed in accordance with one embodiment of the present invention.
5A-5B illustrate examples of two third deep belief network models employed in accordance with another embodiment of the present invention.
Fig. 6 shows a block diagram of an electrical load prediction apparatus according to an embodiment of the present invention.
Fig. 7 shows a block diagram of an electrical load prediction apparatus according to another embodiment of the present invention.
Fig. 8 is a block diagram showing an electric load prediction apparatus according to an embodiment of the present invention.
Detailed Description
Hereinafter, various embodiments of the present invention will be described in detail with reference to the accompanying drawings.
Fig. 1 shows a flow chart of an electrical load prediction method 1 according to an embodiment of the invention. The power load prediction method can be used for predicting the power load, namely the available power amount, of a certain area (such as the east China area, the Hai lake area and a certain cell) or a certain unit in a certain future time (such as one year and one month) by a power company according to the existing historical data, so that power scheduling and planning are performed according to the prediction result.
At step S1, a first deep belief network model for predicting the electrical load from the historical data of the electrical load is trained with the historical data of the electrical load, thereby generating a first prediction result for predicting a future electrical load.
The electric load data is data of the amount of electricity used. The historical data of the electric load is the amount of electricity used per unit time (year, month, day) in the history. For example, assuming that it is currently 12 months in 2014, when it is desired to predict the electricity usage of the cell per month in 2015 years 1-12, the electricity usage of the cell per month in 2014 1-12, the electricity usage per month in 2013 years 1-12, and the electricity usage … … per month in 2012 years 1-12 are historical data for predicting the power load of the cell per month in 2015 years 1-12.
The deep belief network model is one of the known deep learning architectures. As shown in fig. 3, it includes a visible layer 41 and a hidden layer 42. The visible layer 41 is at the lowest layer of the deep belief network model (layer v in fig. 3. the hidden layer may be multiple layers (three layers h in fig. 3) 1 、h 2 、h 3 )。
The deep belief network model is characterized in that the characteristics of data can be extracted from the data, so that the relationship behind the data can be found. Therefore, the layers in fig. 3 can be considered as a layer of process for extracting features of data and finding hidden relationships behind the data from bottom to top. The features are vectors of a certain dimension. Typically, a node in the visible layer 41 receives a plurality of input data. Each node of the first hidden layer h1 extracts each input data from each input data of the visible layer 41A feature of an aspect. Thus, the first hidden layer h 1 Each node of (a) has a relationship with each node (input data) of the visible layer 41 (see the constrained boltzmann function model). Similarly, the second hidden layer h 2 Respectively from the first hidden layer h 1 Further extracting features of a certain aspect from each input data. Thus, the second hidden layer h 2 Each node of (a) and the first hidden layer h 1 A relationship occurs with each node of. Similarly, the third hidden layer h 3 Each node of (a) and the second hidden layer h 2 A relationship occurs with each node of. And regarding the relation between the specific node at the upper layer and the specific node at the lower layer in the deep belief network model as the coefficient of the deep belief network model. The coefficients are generated during model training by cognitive weights and generation weights. The deep belief network model can be viewed as a coefficient matrix that includes many coefficients.
The process of constructing the deep belief network model actually comprises two processes: the former process is a process of extracting features from input data layer by layer to obtain the relationship between each node of the previous layer and each node of the next layer, and obtaining each coefficient in a coefficient matrix; the latter process is a process of verifying each coefficient in the coefficient matrix from top to bottom from the deep belief network model by backward fitting. Deep belief network model in an embodiment of the present invention, for example, when it is required to predict electricity consumption of a certain cell 2014 each day from 12 months to 4 days to 6 days, assuming that electricity consumption L1, L2 and L3 of the cell 2014 each day from 12 months to 1 days to 3 days are historically stored, electricity consumption L1, L2 and L3 of the cell 2014 each day from 12 months to 1 days to 3 days can be respectively input into the first deep belief network model N1 as respective input data, and a first prediction result Y1 for predicting electricity consumption of the cell 2014 each day from 12 months to 4 days to 6 days is generated, as shown in fig. 4. Of course, the above are merely examples. In practice, in order to improve the accuracy of the prediction result, the used amount of electricity for a longer time is generally used as the history data.
The deep belief network model may have different building algorithms. In addition to the above-described Boltzmann function-limiting model, methods such as sparse coding, convolutional neural network, and the like can be employed.
In addition, in the case where the history data of the electric load lacks a part, the missing part may be predicted from the electric load data before the missing part by the first deep belief network model and completed.
For example, it is assumed that only the electricity consumption of the cell 2014 on 12 months and 1-2 days is stored, and the electricity consumption of 3 days is not stored. At this time, only the electricity consumption of the cell 2014 in 12 months and 1-2 days can be input into the first deep belief network model. The first deep belief network model predicts the electricity consumption of the cell 2014 in 12 months and 3 days according to the electricity consumption of the cell 2014 in 12 months and 1-2 days, and then completes the electricity consumption by using the prediction result. And inputting the supplemented electricity consumption of 12 months and 1-3 days in 2014 into a first deep belief network model, and outputting a first prediction result Y1 for predicting the electricity consumption of the cell 2014 in 12 months and 4-6 days. In this way, the problem of how to improve the accuracy of predicting the electrical load in the case where the historical data is partially missing is further solved.
At step S3, a second deep belief network model that predicts the electrical load from the historical data of the electrical load and the historical data of the predefined dominant impact factor is trained with the historical data of the electrical load and the historical data of the predefined dominant impact factor to generate a second prediction result for predicting the future electrical load.
The dominant influence factor refers to a main factor that affects the electrical load. The dominant impact factors for the electrical load may be many and often cannot be met in practice, and some predefined dominant impact factors, i.e. predefined dominant impact factors, are usually adopted. For example, it includes at least one of weather data, temperature data, human context data, and the like. The weather data refers to data related to weather conditions such as sunny, cloudy, and cloudy. For example, in the time of the weather, people tend to prefer to stay at home, and therefore, the use of electricity may be generally used. Temperature data refers to data relating to temperature. For example, in cold winter and hot summer, electricity is often used more by turning on the air conditioner. The human environment data is data reflecting social human factors that may affect the power load in addition to natural environments such as weather, temperature, and the like, for example, recently, a coal system issues a reward measure for a small monthly power consumption of workers, which is a coal system worker cell that may be used later with less power.
Since the predefined dominant impact factor includes weather data, temperature data, human environmental data, etc., the historical data of the predefined dominant impact factor may include historical data of weather data, temperature data, human environmental data, etc. Still taking the prediction of the electricity consumption of a certain cell 2014 at 12 months and 4-6 days as an example, as shown in fig. 4, besides the electricity consumption of 2014 at 12 months and 1-3 days, the prediction is also performed by using the weather data and the temperature data of 2014 at 12 months and 1-3 days. The weather data of 12 months, 1-3 days in 2014 are W1, W2 and W3 respectively. The temperature data of 12 months, 1-3 days in 2014 are T1, T2 and T3 respectively. In this way, the data input to the nodes of the visible layer of the second deep belief network model N4 are not one-dimensional data, but three-dimensional vectors, which are (L1, W1, T1), (L2, W2, T2), (L3, W3, T3), respectively. The output of the second deep belief network model N4 is Y4, the second prediction result. Unlike the first prediction result Y1, which focuses on examining the trend of the change in the power load data itself, the second prediction result Y4 is a power load prediction result which focuses on examining the influence of the predefined dominant influence factor on the power load data.
The deep belief network model can not only process one-dimensional input, namely input historical data a 1 、a 2 ……a n (n is a natural number indicating the number of input history data), the future data a is predicted from the input history data n+1 、a n+2 … …, and can also process multi-dimensional input, i.e. in accepting historical data a 1 、a 2 ……a n While receiving the constraint b 1 、b 2 ……b n ,c 1 、c 2 ……c n And the like, and a is learned through layer-by-layer feature extraction 1 、a 2 ……a n And b 1 、b 2 ……b n ,c 1 、c 2 ……c n Etc., output in consideration of the constraint b 1 、b 2 ……b n ,c 1 、c 2 ……c n Equal pair a 1 、a 2 ……a n Predicted future data a of influence of n+1 、a n+2 … … are provided. The inventor of the invention applies the characteristic of the deep belief network model in the invention, not only a first prediction result of predicting future electric load according to the change trend of historical data of the electric load is generated by using the first deep belief network model, but also a second prediction result of investigating the influence of various predefined dominant influence factors (such as weather data, temperature data and the like) on the electric load is generated by using the second deep belief network model, and a more objective prediction result of not only considering the change trend of the electric load but also independently investigating the electric load and the influence of the surrounding environment can be generated by combining the first prediction result and the second prediction result.
In step S4, a future electrical load is predicted based on the first prediction result and the second prediction result.
In one embodiment, the first prediction result Y1 and the second prediction result Y4 may be simply averaged to obtain the predicted future power load Y ═ (Y1+ Y4)/2.
In another embodiment, the predicted future power load may be obtained by a weighted average method, i.e., Y α Y1+ β Y4, where α + β is 1 and α and β may be set by empirical values, or α and β may be adjusted continuously according to the satisfaction of the final prediction.
Of course, the predicted future electrical load Y may also be derived from the first prediction Y1 and the second prediction Y4 by other means known in the art.
Fig. 2 shows a flow chart of an electrical load prediction method according to another embodiment of the invention. Compared with fig. 1, this embodiment adds step S2: and training a third deep belief network model for predicting and complementing the missing part according to the predefined dominant influence factor before the missing part in the case that the historical data of the predefined dominant influence factor lacks part, by using the historical data of the predefined dominant influence factor, so as to complement the historical data of the predefined dominant influence factor in the case that the historical data of the predefined dominant influence factor lacks part.
The method still takes the example of predicting the electricity consumption of 12 months and 4-6 days in 2014 of a certain cell by using the electricity consumption of 12 months and 1-3 days in 2014 of the cell, weather data and temperature data. It is assumed that only the weather data of the cell 2014 on 12 months and 1-2 days are stored, and the weather data of 3 days are missing. At this time, as shown in fig. 5A, only the weather data of the cell 2014 each day on 12 months 1-2 days may be first input into the third deep belief network model N2. The third deep belief network model N2 predicts weather data of 12 months and 3 days in the cell 2014 according to the weather data of 12 months and 1-2 days in the cell 2014, and then complemental missing data is inversely calculated by using a coefficient matrix of the third deep belief network model N2 obtained by training. The supplemented weather data for days 1-3 of 12 months and 2014 is used by the second deep belief network model N4 in the prediction of step S3. Y2 is a prediction result about future weather, but is not used in the embodiment of the present invention, because the embodiment of the present invention uses the third deep belief network model N2 mainly to complete weather data when some historical data of weather is missing, and does not use the predicted future weather of the third deep belief network model N2.
Similarly, it is assumed that only the temperature data of the cell 2014 on 12 months and 1-2 days are stored, and the temperature data of 3 days are missing. At this time, as shown in fig. 5B, only the temperature data of the cell 2014, 12 months, 1-2 days, may be input into the third deep belief network model N3. And predicting the temperature data of 12, month and 3 days in the cell 2014 by the third deep belief network model N3 according to the temperature data of the cell 2014 in 12, month and 1-2 days, and then inversely calculating the completion missing data by using the coefficient matrix of the third deep belief network model N3 obtained by training. The temperature data for each of days 1-3 of 12 months after completion 2014 was used by the second deep belief network model N4 in the prediction of step S3. Y3 is a predicted result regarding the future temperature, but is not used in the embodiment of the present invention, because the embodiment of the present invention uses the third deep belief network model N2 mainly for completing the temperature data when the historical data of the partial temperature is missing, and does not use the future temperature predicted by the third deep belief network model N2.
In this way, the problem of how to improve the accuracy of predicting the electrical load in case of missing parts of the historical data of the predefined dominant impact factor (e.g. missing parts of the historical data of weather, missing parts of the historical data of temperature) is further solved.
Preferably, in the above process, the coefficient values of the first, second and third depth belief network models are obtained by greedy unsupervised feature learning. That is, feature extraction of different layers of the first, second and third depth belief network models is based on greedy unsupervised feature learning. This makes feature extraction from the next layer to the previous layer very natural without human intervention. Feature extraction at different layers of some deep belief network models requires human correction of the extraction results. Since greedy unsupervised feature learning can clearly express hidden features and complex patterns of historical data and data correlation factors, when part of the historical data or part of the historical data of the data correlation factors is missing, it is easy to correct the hidden relationships among the data by finding the hidden relationships among the data. Therefore, the nature of the load pattern can be restored even in the absence of history data.
Preferably, the normalized historical data of the power load or the normalized historical data of the predefined dominant impact factor is used when training the first deep belief network model with the historical data of the power load, or training the second deep belief network model with the historical data of the power load and the historical data of the predefined dominant impact factor, or training the third deep belief network model with the historical data of the predefined dominant impact factor. The normalization refers to unifying the power load data or the predefined dominant influence factors in different units and different ranges into the power load data or the predefined dominant influence factors in the same unit and the same range. The range here may be a range of regions, a range of statistical periods, or other various ranges. As an example of normalization in units, the stored temperature data may be expressed in degrees celsius, and some may be expressed in degrees fahrenheit, which may be uniformly adjusted to be expressed in degrees celsius. As the normalization of the region range, the stored electricity consumption in 12 months and 1 days in 2014 is the electricity consumption of the whole cell, the stored electricity consumption in 12 months and 2 days in 2014 is the electricity consumption of a half cell (for example, the whole cell has one of two buildings), and the electricity consumption in 12 months and 2 days in 2014 can be multiplied by 2 to be reduced into the electricity consumption of the whole cell, so that the electricity consumption has the same comparison basis. As an example of the normalization of the range of the statistical period, the stored electricity consumption of 12 months and 1 day in 2014 is the electricity consumption of 24 hours in the whole cell, the stored electricity consumption of 12 months and 2 days in 2014 is the electricity consumption of 12 hours in the whole cell, and the electricity consumption of 24 hours in the whole cell can be estimated by multiplying the electricity consumption of 12 months and 2 days in 2014 by a certain empirical coefficient (not necessarily 2 due to unbalanced electricity consumption in daytime and evening), so that the electricity consumption has the same comparison basis. Through the normalization, the prediction precision is also improved.
As shown in fig. 6, another embodiment of the present invention provides an electrical load prediction apparatus 2, which includes a first training unit 21, a second training unit 23, and a prediction unit 24. The first training unit 21 is configured to train a first deep belief network model for predicting the electrical load from the historical data of the electrical load with the historical data of the electrical load, thereby generating a first prediction result for predicting a future electrical load. The second training unit 23 is configured to train a second deep belief network model predicting the electrical load from the historical data of the electrical load and the historical data of the predefined dominant impact factor with the historical data of the electrical load and the historical data of the predefined dominant impact factor, thereby generating a second prediction result for predicting the future electrical load. The prediction unit 24 is configured to predict a future electrical load from the first prediction result and the second prediction result. The units in fig. 6 may be implemented by software, hardware (e.g., integrated circuit, FPGA, etc.), or a combination of software and hardware. Predefining dominant impact factors
Optionally, the first training unit 21 is further configured to predict the missing part from the power load data before the missing part with the first deep belief network model and to complement the historical data of the power load in case the historical data of the power load is partially missing.
Alternatively, as shown in fig. 7, the electrical load prediction apparatus 2 may further include: a third training unit 22 configured to train a third deep belief network model for predicting the missing part from the predefined dominant influence before the missing part and complementing the historical data of the predefined dominant influence if the historical data of the predefined dominant influence is partially missing with the historical data of the predefined dominant influence, thereby complementing the historical data of the predefined dominant influence if the historical data of the predefined dominant influence is partially missing.
Optionally, the coefficient values of the first, second and third depth belief network models are obtained by greedy unsupervised feature learning.
Optionally, the historical data of the electrical load is normalized historical data of the electrical load, and the historical data of the predefined dominant impact factor is normalized historical data of the predefined dominant impact factor.
Optionally, the predefined dominant impact factor further comprises at least one of weather data, temperature data, human context data.
Referring now to fig. 8, there is shown a schematic diagram of an electrical load prediction apparatus 3 in accordance with one embodiment of the present invention. As shown in fig. 8, the electrical load prediction apparatus 3 may include a memory 31 and a processor 32. The memory 31 may store executable instructions. The processor 32 may implement the operations performed by the various units of the apparatus 2 described above according to executable instructions stored by the memory 31.
Additionally, embodiments of the present invention also provide a machine-readable medium having stored thereon executable instructions that, when executed, cause a machine to perform operations performed by processor 32.
It will be understood by those skilled in the art that various changes and modifications may be made to the embodiments disclosed above without departing from the spirit of the invention. Accordingly, the scope of the invention should be determined from the following claims.
Claims (8)
1. An electrical load prediction method (1) comprising:
training a first deep belief network model for predicting the electrical load from the historical data of the electrical load with the historical data of the electrical load, thereby producing a first prediction result for predicting a future electrical load; in the case where the history data of the electric load is partially missing, predicting the missing part from the electric load data before the missing part by the first deep belief network model and complementing the history data of the electric load (S1);
training a second deep belief network model predicting the electrical load from the historical data of the electrical load and the historical data of the predefined dominant impact factor with the historical data of the electrical load and the historical data of the predefined dominant impact factor, thereby generating a second prediction result for predicting a future electrical load (S3); and
predicting a future electrical load based on the first prediction result and the second prediction result (S4);
the method further comprises the following steps: training a third deep belief network model for predicting the missing part from the predefined dominant influence before the missing part and complementing the historical data of the predefined dominant influence in case the historical data of the predefined dominant influence is missing part with the historical data of the predefined dominant influence, thereby complementing the historical data of the predefined dominant influence in case the historical data of the predefined dominant influence is partially missing (S2).
2. The electrical load prediction method (1) according to claim 1, wherein the coefficient values of the first, second and third depth belief network models are learned using a greedy unsupervised feature.
3. The electrical load prediction method (1) as claimed in claim 1, wherein the historical data of the electrical load is normalized historical data of the electrical load and the historical data of the predefined dominant impact factor is normalized historical data of the predefined dominant impact factor.
4. The electrical load prediction method (1) according to claim 1, wherein the predefined dominant impact factor comprises at least one of weather data, temperature data, human environment data.
5. An electrical load prediction device (2) comprising:
a first training unit (21) configured to train a first deep belief network model for predicting the electrical load from the historical data of the electrical load with the historical data of the electrical load, thereby generating a first prediction result for predicting a future electrical load;
a second training unit (23) configured to train a second deep belief network model predicting the electrical load from the historical data of the electrical load and the historical data of the predefined dominant impact factor with the historical data of the electrical load and the historical data of the predefined dominant impact factor, thereby generating a second prediction result for predicting the future electrical load;
a prediction unit (24) configured to predict a future electrical load from the first prediction result and the second prediction result;
wherein the first training unit (21) is further configured to predict the missing part from power load data preceding the missing part with the first deep belief network model and to complement the historical data of the power load in case the historical data of the power load is partially missing;
the device also includes: a third training unit (22) configured to train a third deep belief network model for predicting the missing part from the predefined dominant influence before the missing part and complementing the historical data of the predefined dominant influence if the historical data of the predefined dominant influence is partially missing with the historical data of the predefined dominant influence, thereby complementing the historical data of the predefined dominant influence if the historical data of the predefined dominant influence is partially missing.
6. The electrical load prediction apparatus (2) according to claim 5, wherein the coefficient values of the first, second and third depth belief network models are learned using a greedy unsupervised feature.
7. The electrical load prediction device (2) according to claim 5, wherein the historical data of the electrical load is the historical data of the normalized electrical load and the historical data of the predefined dominant impact factor is the historical data of the normalized predefined dominant impact factor.
8. The electrical load prediction apparatus (2) according to claim 5, wherein the predefined dominant impact factor comprises at least one of weather data, temperature data, human context data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410838273.3A CN105809264B (en) | 2014-12-29 | 2014-12-29 | Power load prediction method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410838273.3A CN105809264B (en) | 2014-12-29 | 2014-12-29 | Power load prediction method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105809264A CN105809264A (en) | 2016-07-27 |
| CN105809264B true CN105809264B (en) | 2022-08-02 |
Family
ID=56979890
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410838273.3A Active CN105809264B (en) | 2014-12-29 | 2014-12-29 | Power load prediction method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105809264B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106529747A (en) * | 2017-01-04 | 2017-03-22 | 成都四方伟业软件股份有限公司 | Power load predicting method and system based on large data |
| KR101917729B1 (en) * | 2017-03-06 | 2018-11-12 | 온동네피엠씨 주식회사 | Average power consumption control system |
| CN109903165B (en) * | 2018-12-14 | 2020-10-16 | 阿里巴巴集团控股有限公司 | Model merging method and device |
| CN110212520A (en) * | 2019-05-24 | 2019-09-06 | 国网天津市电力公司 | A kind of power predicating method based on convolutional neural networks |
| CN112381266B (en) * | 2020-10-22 | 2024-01-09 | 国网湖北省电力有限公司武汉供电公司 | System and method for predicting future power supply quantity based on historical power supply and weather data |
| CN117353300B (en) * | 2023-12-04 | 2024-02-23 | 拓锐科技有限公司 | Rural power consumption demand analysis method based on big data |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103514491A (en) * | 2013-10-18 | 2014-01-15 | 国网四川省电力公司自贡供电公司 | Power load forecasting method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2004094437A (en) * | 2002-08-30 | 2004-03-25 | Fuji Electric Holdings Co Ltd | Data prediction method and data prediction system |
-
2014
- 2014-12-29 CN CN201410838273.3A patent/CN105809264B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103514491A (en) * | 2013-10-18 | 2014-01-15 | 国网四川省电力公司自贡供电公司 | Power load forecasting method |
Non-Patent Citations (2)
| Title |
|---|
| 人工神经网络在上海电力负荷预测中的应用;张国忠等;《华东电力》;20020228(第2期);第7-9页 * |
| 基于深度信念网络的短期电力负荷预测;肖同录等;《电子科技》;20140619(第10期);第186-187页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105809264A (en) | 2016-07-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105809264B (en) | Power load prediction method and device | |
| JP6384065B2 (en) | Information processing apparatus, learning method, and program | |
| Xu et al. | Discrete time–cost–environment trade-off problem for large-scale construction systems with multiple modes under fuzzy uncertainty and its application to Jinping-II Hydroelectric Project | |
| Wan et al. | Probabilistic forecasting of wind power generation using extreme learning machine | |
| Zhou et al. | Remaining useful life prediction of individual units subject to hard failure | |
| JP2009294969A (en) | Demand forecast method and demand forecast device | |
| CN102930155B (en) | Obtain the method and device of the early-warning parameters of electricity needs | |
| JP6498976B2 (en) | Estimation apparatus, estimation method, and computer program | |
| JP2014157457A (en) | Prediction device and prediction method | |
| Ahmadi et al. | A lexicographic optimization and augmented ϵ-constraint technique for short-term environmental/economic combined heat and power scheduling | |
| JP2019087030A (en) | Prediction model generation device, prediction model generation method and prediction model generation program | |
| JP6086875B2 (en) | Power generation amount prediction device and power generation amount prediction method | |
| JP2019179538A (en) | Prediction method for predicting construction price of building and prediction method for predicting its construction period | |
| CN113222403A (en) | Power adjusting method and device based on big data, storage medium and electronic equipment | |
| CN109409561A (en) | The construction method of Multiple Time Scales time series collaborative forecasting model | |
| Ruan et al. | Estimating demand flexibility using Siamese LSTM neural networks | |
| El Kontar et al. | Profiling occupancy patterns to calibrate Urban Building Energy Models (UBEMs) using measured data clustering | |
| CN103106331A (en) | Photo-etching line width intelligence forecasting method based on dimension-reduction and quantity-increment-type extreme learning machine | |
| CN103885867A (en) | Online evaluation method of performance of analog circuit | |
| Qadrdan et al. | Probabilistic wind power forecasting using a single forecast | |
| Markoska et al. | Towards smart buildings performance testing as a service | |
| CN110007371A (en) | Wind speed forecasting method and device | |
| CN105894138A (en) | Optimum weighted composite prediction method for shipment amount of manufacturing industry | |
| CN110753366A (en) | Prediction processing method and device for industry short message gateway capacity | |
| Wilson et al. | Use of meteorological data for improved estimation of risk in capacity adequacy studies |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |