Multi-channel power load prediction device and method based on optimal data selection

Abstract

The invention provides a multichannel power load prediction device and method based on optimal data selection. The device comprises a data collection module used for collecting and formatting data of a power load prediction scene; the optimal similar day section selection module is used for calculating an optimal historical data period according to the data formatted by the data collection module and extracting historical data of the optimal historical data period as sub-data; the feature processing module is used for adding new data features to the data at each time point according to the sub-data extracted by the optimal similar day section selection module; and the multi-channel model module is used for obtaining a prediction result according to the data obtained after the new data features are added by the feature processing module.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to a multi-channel electric power load forecasting device and method based on optimal data selection. Background Art

[0002] With the deepening of the national power market reform, power load forecasting plays an important role in the power market. Accurate power load forecasting can not only help power dispatching agencies plan unit dispatching, line maintenance and other tasks, but also help various market players participating in power market transactions to formulate better trading strategies, thereby promoting the entire power market reform towards a benign and rapid development direction.

[0003] In general, the power load forecasting process can predict the power load data for the next 24 hours, 48 ​​hours or even longer by constructing a machine learning or deep learning algorithm model and using historical data for training. However, in the current process of power load forecasting, people often choose to train and predict all historical data of all data features used in the target task in a machine learning model or a deep learning model. Although there is no problem with such a data training method, especially when the data volume is large, the data changes are small and the data is relatively stable, such a data training method is more likely to ignore the local characteristics of the newer data in the time period, or the similarity of recent data, especially the climate instability caused by the current global warming, or the changes in production demand in the short term, etc. In this case, the prediction of these data is easily affected by the length of the historical data cycle used. Therefore, based on the original invention patent, the present invention proposes an optimization method for selecting historical data to deal with this problem.

[0004] In addition, when predicting power load, we may design more complex machine learning models or deep learning models. One of the main technical features of these models is to capture some features in the data, such as peak features, shape features, etc. In the process of model design, most of the original patents choose to use one or several models to form a single model channel for model training. This will place high demands on whether the model used can capture enough data features. If the model used is not capable enough, it will greatly affect the final power load prediction effect.

[0005] The patent with publication number CN115730717A proposes a multi-channel network design method based on CNN+BiLSTM, but its network design mainly includes four channels, and each channel uses a network with a CNN+BiLSTM structure, which lacks flexibility and may not be able to accurately capture different features in the data.

[0006] The patent with publication number CN108229754A proposes a short-term load forecasting method based on segmented similar days and LM-BP network. The method first uses 5 meteorological features to describe daily weather, and takes all data on the same day of each week in historical data except holidays as candidate data, and then considers the trend data of the previous k days of the selected historical daily data and the trend data of the previous k days before the forecast day for similarity determination. Although this method considers several factors that affect the prediction accuracy of similar days to a large extent, such as weather, trend data of the previous k days, etc., its trend data of the previous k days considers the average trend of 1 to k days, and the selection of k value is easily restricted by artificial selection. Thereby increasing the training time cost of the entire model, and the average trend of the data may not reflect the real change of the data, especially in the scenario of wind power generation prediction where the data fluctuation is particularly drastic, the average trend is not effective, and the meteorological data of the same day of each week in the historical data cannot be used as an effective reference. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a multi-channel power load forecasting device and method based on optimal data selection in view of the above-mentioned defects in the prior art, which can optimize the calculation of similar day data that adapts to multiple scenarios and situations where data changes are more obvious.

[0008] According to the present invention, a multi-channel power load forecasting device based on optimal data selection is provided, which comprises: A data collection module is used to collect and format data for power load forecasting scenarios; The optimal similar day segment selection module is used to calculate the optimal historical data period according to the data formatted by the data collection module, and extract the historical data of the optimal historical data period as sub-data; The feature processing module is used to add new data features to the data at each time point for the sub-data extracted by the optimal similar day segment selection module; The multi-channel model module is used to obtain prediction results based on the data obtained after adding new data features to the feature processing module.

[0009] Preferably, the multi-channel model module includes a first main channel and a second main channel, the data input of the first main channel and the second main channel is the data obtained after the feature processing module adds new data features; the first main channel only contains a single model, and the first main channel obtains an intermediate result by calculating the data obtained after the feature processing module adds new data features through a single model; the second main channel includes multiple independent sub-channels with different models, and each independent sub-channel obtains an intermediate result by calculating the data obtained after the feature processing module adds new data features through its own model, and all independent sub-channels perform nonlinear processing through nonlinear activation function operations to obtain the intermediate result finally generated by the second main channel, and the intermediate result finally generated by the second main channel is added to the intermediate result output by the first main channel, and the output result is obtained by linear transformation through a fully connected layer.

[0010] Preferably, the second main channel contains 4 single models; the first sub-channel is the CNN+ATTENTION channel, which calculates each data feature sequence in the data and outputs the intermediate result obtained by the channel model operation; the second sub-channel is the WNN+Shaplet ATTENTION channel, which captures the shape features of each data feature sequence and outputs the intermediate result generated by the model; the third sub-channel is the Transformer Encoder channel, which is used to capture the data correlation or association between different data features at each time point, and outputs the intermediate result generated by the model; the fourth sub-channel is the CNN+GRU channel, which extracts data features through the CNN network, calculates the residual between vectors, further extracts the residual data features through the GRU network, and outputs the intermediate result.

[0011] Preferably, the single model is a machine learning model among XGBoost, CatBoost, and LightGBM.

[0012] Preferably, the optimal similar day segment selection module determines the time point when the change of power load compared to the current time point is greater than a threshold as the target day, and traverses the data of a predetermined time period before the target day to calculate the similar day prediction data, and compares the calculated similar day prediction data with the actual data of the target day, and takes the time period with the smallest RMSE as the optimal historical data period.

[0013] Preferably, the new data features include a first new data feature and a second new data feature; wherein the first new data feature is calculated by the difference between the data at the current time point and the data at the previous time point, and is used to characterize the data characteristics of the direction and magnitude of data change; the second new data feature is composed of the average value of the data at the current time point and the data at the time point corresponding to 24 hours before the current time point, and is mainly used to reflect the data change trend.

[0014] According to the present invention, there is also provided a multi-channel power load forecasting method based on optimal data selection, comprising: The data collection module is used to collect and format the data of the power load forecasting scenario; Using the optimal similar day segment selection module, the optimal historical data period is calculated according to the data formatted by the data collection module, and the historical data of the optimal historical data period is extracted as sub-data; Using the feature processing module, new data features are added to the data at each time point for the sub-data extracted by the optimal similar day segment selection module; The multi-channel model module is used to obtain the prediction results based on the data obtained after adding new data features from the feature processing module.

[0015] Preferably, the multi-channel model module includes a first main channel and a second main channel, the data input of the first main channel and the second main channel is the data obtained after the feature processing module adds new data features; the first main channel only contains a single model, and the first main channel obtains an intermediate result by calculating the data obtained after the feature processing module adds new data features through a single model; the second main channel includes multiple independent sub-channels with different models, and each independent sub-channel obtains an intermediate result by calculating the data obtained after the feature processing module adds new data features through its own model, and all independent sub-channels perform nonlinear processing through nonlinear activation function operations to obtain the intermediate result finally generated by the second main channel, and the intermediate result finally generated by the second main channel is added to the intermediate result output by the first main channel, and the output result is obtained by linear transformation through a fully connected layer.

[0016] Preferably, the second main channel contains 4 single models; the first sub-channel is the CNN+ATTENTION channel, which calculates each data feature sequence in the data and outputs the intermediate result obtained by the channel model operation; the second sub-channel is the WNN+Shaplet ATTENTION channel, which captures the shape features of each data feature sequence and outputs the intermediate result generated by the model; the third sub-channel is the Transformer Encoder channel, which is used to capture the data correlation or association between different data features at each time point, and outputs the intermediate result generated by the model; the fourth sub-channel is the CNN+GRU channel, which extracts data features through the CNN network, calculates the residual between vectors, further extracts the residual data features through the GRU network, and outputs the intermediate result.

[0017] Preferably, the single model is a machine learning model among XGBoost, CatBoost, and LightGBM. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] A more complete understanding of the present invention and its attendant advantages and features will be more readily appreciated by reference to the following detailed description taken in conjunction with the accompanying drawings, in which:

[0019] Figure 1 A functional block diagram of a multi-channel power load prediction device based on optimal data selection according to a preferred embodiment of the present invention is schematically shown.

[0020] Figure 2 A schematic diagram of a multi-channel model module of a multi-channel power load prediction device based on optimal data selection according to a preferred embodiment of the present invention is schematically shown.

[0021] It should be noted that the drawings are used to illustrate the present invention, rather than to limit the present invention. Note that the drawings showing the structures may not be drawn to scale. In addition, in the drawings, the same or similar elements are marked with the same or similar reference numerals. DETAILED DESCRIPTION

[0022] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0024] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0025] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0026] Similar day data calculation is of great significance in the power load forecasting scenario, but since power load forecasting has many usage scenarios and faces different forecast types, such as power grid load forecasting, photovoltaic output load forecasting, wind output load forecasting and many other scenarios; when historical data is insufficient and data fluctuations are unstable, similar day data calculation may lose its role as an important tool for data forecasting. To this end, the present invention proposes an optimization device and method for similar day data calculation that is adaptable to multiple scenarios and to situations where data changes are more obvious.

[0027] Moreover, the present invention also proposes a network optimization model that uses a multi-channel machine learning and deep learning model in a mixed manner. The model uses a variety of models or model combinations to form different model channels, thereby improving the model's ability to capture data features and thus improving the accuracy of model prediction. At the same time, the multi-channel design can use multiple GPUs for parallel training, thereby reducing the time cost and resource utilization of model training.

[0028] Figure 1 The functional block diagram of a multi-channel power load forecasting device based on optimal data selection according to a preferred embodiment of the present invention is schematically shown. Figure 1 As shown, a multi-channel power load forecasting device based on optimal data selection according to a preferred embodiment of the present invention includes: Data collection module 10: collects and formats data of power load forecasting scenarios.

[0029] Specifically, the data collection module is mainly used for data collection and formatting corresponding to the power load forecasting scenario. Here, the power load data includes but is not limited to regional power load data, new energy load data, meteorological data, calendar data, electricity price policy data, industrial chain data, etc. The formatting process is mainly used to organize the original data into a two-dimensional table.

[0030] The optimal similar day segment selection module 20 is used to calculate the optimal historical data period according to the data formatted by the data collection module, and extract the historical data of the optimal historical data period as sub-data.

[0031] Preferably, the optimal similar day segment selection module 20 determines the time point when the change of power load compared to the current time point is greater than a threshold as the target day, and traverses the data of a predetermined time period before the target day to calculate the similar day prediction data, and compares the calculated similar day prediction data with the actual data of the target day, and takes the time period with the smallest RMSE as the optimal historical data period.

[0032] For example, first, the module analyzes formatted data, such as industry and grid-related data (optionally, policy data), to derive the time point at which a significant change occurred in history compared to the current time point. For example, for the photovoltaic output load forecast, the time point at which a significant change occurred in history is the time when the last new photovoltaic power station was publicly announced to be connected to the grid.

[0033] For example, if the time period from this time point to the present is less than 3 months, then set the step size to 1 month, traverse the data of the past year, calculate the forecast data of similar days, compare it with the actual data of the target day, calculate the RMSE (root mean square error) indicator, and take the time period with the smallest RMSE as the target historical data period. The target day here is the data of the last day corresponding to the historical data, not the target forecast day.

[0034] Based on the time period selected above, the similarity prediction data of the target prediction day is calculated, and the historical data corresponding to the time period is extracted (which can be saved separately) as sub-data.

[0035] The feature processing module 30 is used to add new data features to the data at each time point for the sub-data extracted by the optimal similar day segment selection module.

[0036] Preferably, the new data features include a first new data feature and a second new data feature; wherein the first new data feature is calculated by the difference between the data at the current time point and the data at the previous time point, and is used to characterize the data characteristics of the direction and magnitude of data change; the second new data feature is composed of the average value of the data at the current time point and the data at the time point corresponding to 24 hours before the current time point, and is mainly used to reflect the data change trend.

[0037] The multi-channel model module 40 is used to obtain a prediction result based on the data obtained after the feature processing module 30 adds new data features.

[0038] In a specific embodiment, the multi-channel model module 40 includes a first main channel and a second main channel, and the data input of the first main channel and the second main channel is the data obtained after the feature processing module 30 adds new data features.

[0039] Preferably, the first main channel contains only a single model, which may be a machine learning model such as XGBoost / CatBoost / LightGBM, for example, preferably a CatBoost model.

[0040] The first main channel obtains the intermediate result by calculating the data obtained after the feature processing module 30 adds new data features through a single model.

[0041] The second main channel includes multiple independent sub-channels with different models. Each independent sub-channel obtains an intermediate result by calculating the data obtained after adding new data features to the feature processing module 30 through its own model. All independent sub-channels perform nonlinear processing through nonlinear activation function operation to obtain the intermediate result finally generated by the second main channel. The intermediate result finally generated by the second main channel is added to the intermediate result output by the first main channel, and a linear transformation is performed through the fully connected layer to obtain the output result, which is the required prediction result.

[0042] It should be noted that the present invention does not limit the number of subchannels used here, and can be configured according to specific actual conditions. For example, it is preferred to adopt a configuration of 4 independent channels, wherein the first subchannel is a CNN+ATTENTION channel, which calculates each data feature sequence in the data and outputs the intermediate result obtained by the channel model operation; the second subchannel is a WNN+Shaplet ATTENTION channel, which is mainly used to capture the shape features of each data feature sequence and output the intermediate result generated by the model; the third subchannel is a Transformer Encoder channel, which is mainly used to capture the data correlation or association between different data features at each time point, and output the intermediate result generated by the model; the fourth subchannel is a CNN+GRU channel, which extracts data features through the CNN network, calculates the residual between vectors, further extracts residual data features through the GRU network, and outputs the intermediate result.

[0043] After each sub-channel outputs the intermediate result, the intermediate results generated by all sub-channels are collected into the same GPU through the devices involved. These intermediate results are nonlinearly processed by nonlinear activation function operation to obtain the intermediate result finally generated by the second main channel, and then added to the intermediate result output by the first main channel, and linearly transformed through the fully connected layer to obtain the output result, which is the required prediction result.

[0044] Preferably, the above-mentioned device will load different model channels onto different GPUs to achieve parallel training of the model, thereby improving efficiency.

[0045] In addition, for example, there may be a data normalization process and a denormalization process before and after the execution of the above steps, for example, the RevIN method may be used for normalization and denormalization processing.

[0046] In addition, it should be noted that, unless otherwise specified, the terms "first", "second", "third", etc. in the specification are only used to distinguish the various components, elements, steps, etc. in the specification, and are not used to indicate the logical relationship or sequential relationship between the various components, elements, steps, etc.

[0047] It is to be understood that, although the present invention has been disclosed as a preferred embodiment, the above embodiment is not intended to limit the present invention. For any person skilled in the art, without departing from the scope of the technical solution of the present invention, the technical content disclosed above can be used to make many possible changes and modifications to the technical solution of the present invention, or modified into equivalent embodiments of equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A multi-channel power load forecasting device based on optimal data selection, characterized in that include: A data collection module is used to collect and format data for power load forecasting scenarios; The optimal similar day segment selection module is used to calculate the optimal historical data period according to the data formatted by the data collection module, and extract the historical data of the optimal historical data period as sub-data; The feature processing module is used to add new data features to the data at each time point for the sub-data extracted by the optimal similar day segment selection module; The multi-channel model module is used to obtain prediction results based on the data obtained after adding new data features to the feature processing module.

2. The multi-channel power load forecasting device based on optimal data selection according to claim 1 is characterized in that: The multi-channel model module includes a first main channel and a second main channel. The data input of the first main channel and the second main channel is the data obtained after the feature processing module adds new data features. The first main channel only contains a single model. The first main channel obtains the intermediate result by calculating the data obtained after the feature processing module adds new data features through a single model. The second main channel includes multiple independent sub-channels with different models. Each independent sub-channel obtains the intermediate result by calculating the data obtained after the feature processing module adds new data features through its own model. All independent sub-channels perform nonlinear processing through nonlinear activation function operations to obtain the intermediate result finally generated by the second main channel. The intermediate result finally generated by the second main channel is added to the intermediate result output by the first main channel, and the output result is obtained by linear transformation through a fully connected layer.

3. The multi-channel power load forecasting device based on optimal data selection according to claim 2 is characterized in that: The second main channel contains 4 single models; The first sub-channel is the CNN+ATTENTION channel, which calculates each data feature sequence in the data and outputs the intermediate result obtained by the channel model operation; the second sub-channel is the WNN+Shaplet ATTENTION channel, which captures the shape features of each data feature sequence and outputs the intermediate result generated by the model; the third sub-channel is the Transformer Encoder channel, which is used to capture the data correlation or association between different data features at each time point, and outputs the intermediate result generated by the model; the fourth sub-channel is the CNN+GRU channel, which extracts data features through the CNN network, calculates the residual between vectors, further extracts the residual data features through the GRU network, and outputs the intermediate result.

4. The multi-channel power load forecasting device based on optimal data selection according to claim 2 is characterized in that: A single model is a machine learning model in XGBoost, CatBoost, and LightGBM.

5. The multi-channel power load forecasting device based on optimal data selection according to claim 1 or 2, characterized in that: The optimal similar day segment selection module determines the time point when the change of power load compared with the current time point is greater than the threshold as the target day, and traverses the data of the predetermined time period before the target day to calculate the similar day prediction data, and compares the calculated similar day prediction data with the actual data of the target day, and takes the time period with the smallest RMSE as the optimal historical data period.

6. The multi-channel power load forecasting device based on optimal data selection according to claim 1 or 2, characterized in that: The new data feature includes a first new data feature and a second new data feature; The first new data feature is calculated by the difference between the data at the current time point and the data at the previous time point, and is used to characterize the data characteristics of the direction and magnitude of data change; the second new data feature is composed of the average value of the data at the current time point and the data at the corresponding time point 24 hours before the current time point, and is mainly used to reflect the trend of data changes.

7. A multi-channel power load forecasting method based on optimal data selection, characterized in that include: The data collection module is used to collect and format the data of the power load forecasting scenario, wherein the formatting process is used to organize the data into a two-dimensional table form; Using the optimal similar day segment selection module, the optimal historical data period is calculated according to the data formatted by the data collection module, and the historical data of the optimal historical data period is extracted as sub-data; Using the feature processing module, new data features are added to the data at each time point for the sub-data extracted by the optimal similar day segment selection module; Using the multi-channel model module, the prediction results are obtained based on the data obtained after adding new data features in the feature processing module; Among them, the multi-channel model module includes a first main channel and a second main channel, and the data input of the first main channel and the second main channel is the data obtained after the feature processing module adds new data features; the first main channel only contains a single model, and the first main channel obtains the intermediate result by calculating the data obtained after the feature processing module adds new data features through a single model; the second main channel includes multiple independent sub-channels with different models, and each independent sub-channel obtains the intermediate result by calculating the data obtained after the feature processing module adds new data features through its own model, and all independent sub-channels perform nonlinear processing through nonlinear activation function operations to obtain the intermediate result finally generated by the second main channel, and the intermediate result finally generated by the second main channel is added to the intermediate result output by the first main channel, and the output result is obtained by linear transformation through a fully connected layer.

8. The multi-channel power load forecasting method based on optimal data selection according to claim 7 is characterized in that: The second main channel contains 4 single models; The first sub-channel is the CNN+ATTENTION channel, which calculates each data feature sequence in the data and outputs the intermediate result obtained by the channel model operation; the second sub-channel is the WNN+Shaplet ATTENTION channel, which captures the shape features of each data feature sequence and outputs the intermediate result generated by the model; the third sub-channel is the Transformer Encoder channel, which is used to capture the data correlation or association between different data features at each time point, and outputs the intermediate result generated by the model; the fourth sub-channel is the CNN+GRU channel, which extracts data features through the CNN network, calculates the residual between vectors, further extracts the residual data features through the GRU network, and outputs the intermediate result.

9. The multi-channel power load forecasting method based on optimal data selection according to claim 7 is characterized in that: A single model is a machine learning model in XGBoost, CatBoost, and LightGBM.

Images