Regional short-term load prediction method and system based on historical load characteristic optimization
By performing Min-Max normalization and multi-time scale feature extraction on historical power load data, combining with CNN-LSTM model for prediction, and restoring the results, the shortcomings of existing power load prediction methods are solved, and the accuracy and reliability of short-term load prediction are significantly improved.
Patent Information
- Application Number
- CN202411679221.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-22
- Publication Date
- 2025-04-25
AI Technical Summary
The existing power load prediction methods have problems such as insufficient data processing, incomplete feature extraction, single model selection, inaccurate prediction result reduction and poor dynamic adaptability, resulting in limited prediction effect and reliability.
The regional short-term load prediction method based on historical load characteristics is adopted to normalize the load data by Min-Max, and a load characteristic matrix of multi-time scale is constructed, and a CNN-LSTM combination model is used for prediction, and the prediction results are finally restored through the maximum and minimum values of the load data.
It improves the accuracy and reliability of short-term power load prediction, can better capture load change trends, enhances the prediction accuracy and robustness of the model, and improves the ability to adapt to emergencies.
Smart Images
Figure CN119891146A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and system for regional short-term load forecasting, and in particular to a method and system for regional short-term load forecasting based on historical load characteristic optimization. Background Art
[0002] The energy revolution has promoted the leapfrog development of social productivity. Today, electricity, as an energy source with an easily adjustable power curve, is the main energy source for driving the operation of machinery and equipment. Therefore, the supply of electricity is an important part of ensuring the sound development of social economy. For example, winter cold waves and useless virtual mining have caused tight power supply and more electricity demand, and put forward higher power supply requirements for the power grid. However, power generation, power supply and power consumption are carried out at the same time, and it is still difficult to store large amounts of electricity. Therefore, in order to match the level of power generation and power consumption as much as possible and formulate the most reasonable power generation plan, it is necessary to accurately predict the power load of the power grid in order to make a reasonable power production plan.
[0003] For the time classification of power load, there are four categories from short to long: ultra-short-term, short-term, medium-term and long-term. Ultra-short-term forecasts are generally in hours, which helps the power grid to fine-tune the power generation strategy; short-term forecasts are in days, which allows power generation to adjust the power generation strategy according to the week and month; medium-term and long-term forecasts are in months and years, respectively, which are references for long-term power development strategies. The forecast of short-term power load is the basis for the good operation of power generation units, and helps to customize power generation plans and power dispatching plans.
[0004] There are some obvious deficiencies in the existing power load forecasting methods: first, insufficient data processing. Many methods only perform simple normalization processing and fail to make full use of the temporal and spatial characteristics of the data, which makes it difficult for the model to capture deep patterns. Second, feature extraction is not comprehensive. Existing methods usually only consider the characteristics of a single time scale, such as daily average load or weekly average load, lack comprehensive consideration of multi-time scale characteristics, and cannot fully reflect the complexity of load changes. Third, the model selection is single. Many methods only rely on traditional statistical models (such as ARIMA, linear regression) or single machine learning models (such as support vector machines). These models perform poorly when processing complex time series data, especially when dealing with nonlinear and long dependencies. Fourth, the prediction result restoration is inaccurate. Existing methods directly reverse the normalized prediction results during the prediction result restoration process, ignoring the maximum and minimum values during data preprocessing, resulting in large prediction result errors. Finally, the dynamic adaptability is poor. Existing methods cannot adjust the prediction model in time when facing sudden events (such as extreme weather, emergencies, etc.), resulting in inaccurate prediction results. These problems limit the effectiveness and reliability of existing methods in practical applications. Summary of the invention
[0005] Purpose of the invention: The purpose of the present invention is to provide a method and system for regional short-term load forecasting based on historical load characteristics optimization, which can reduce the complexity of forecasting.
[0006] Technical solution: The method for regional short-term load forecasting based on historical load characteristic optimization described in the present invention comprises:
[0007] S1: Collect the load data of the area to be predicted every hour, perform Min-Max normalization on the load data, and record the maximum and minimum values of the load;
[0008] S2: Based on the preprocessed load data, obtain load characteristics and construct a load characteristic matrix;
[0009] S3: Use the load feature matrix to train the CNN-LSTM combined model, input the load features after CNN processing into the LSTM to extract time series features and make predictions;
[0010] S4: Based on the maximum and minimum values of the load data, the prediction results of the CNN-LSTM combined model are restored to obtain the final prediction value.
[0011] Furthermore, the step S2 comprises:
[0012] S2.1: Obtain the hourly regional load sequence K+1 years before the forecast date, and calculate the daily average load sequence K years before the forecast date in units of days, with a sequence length of 365(K+1);
[0013] S2.2: Select the historical daily average load K days before the day to be predicted to form a row vector with a length of K, which is used as the recent load characteristic row vector, denoted as c1;
[0014] S2.3: Select the daily average load on the day before the forecast date and the same day in the previous K-1 weeks as the weekly load characteristic row vector, denoted as c2;
[0015] S2.4: Select the daily average load on the day before the forecast date and the same day of the previous K-1 months as the monthly load characteristic row vector, denoted as c3;
[0016] S2.5: Select the daily average load on the day before the forecast date and the same day in the previous K-1 years as the annual load characteristic row vector, denoted as c4;
[0017] S2.6: Select the historical daily average load sequence of K+1 weeks before the day to be predicted, and calculate the weekly average load on a weekly basis to obtain a load sequence of length K+1. Select the first K values and subtract the last value to form a difference sequence of length K. Perform Min-Max normalization on the obtained difference sequence as the weekly load change feature, recorded as c5, to characterize the impact of human activities on the load on a weekly scale;
[0018] S2.7: Select the historical daily average load sequence of K+1 months before the forecast date, and calculate the monthly average load on a monthly basis to obtain a load sequence of length K+1. Select the first K values and subtract the last value to form a difference sequence of length K. Perform Min-Max normalization on the obtained difference sequence as the monthly load change feature, recorded as c6, to characterize the impact of human activities on the load on a monthly scale;
[0019] S2.8: Select the historical daily average load sequence of K+1 years before the day to be predicted, and calculate the annual average load on an annual basis to obtain a load sequence of length K+1. Select the first K values and subtract the last value to form a difference sequence of length K. Perform Min-Max normalization on the obtained difference sequence as the annual load change feature, recorded as c7, to characterize the impact of human activities on the load on an annual scale;
[0020] S2.9: Stack the features calculated in steps S2.2 to S2.8 by row to obtain a load feature matrix Its dimension is 7×K.
[0021] Furthermore, the step S3 comprises:
[0022] S3.1: Construct a CNN-LSTM model. CNN uses a one-dimensional convolutional neural network. CNN and LSTM are cascaded, and the last time step of the LSTM network is taken as the prediction result.
[0023] S3.2: Train the CNN-LSTM model. The model input is the load feature matrix, the model output is the predicted load, and the training loss is the MSE error.
[0024] Furthermore, the step S4 comprises:
[0025] Suppose the prediction result of the CNN-LSTM combination model is x ′ pre , then the final predicted value x pre for:
[0026] x pre =x ′ pre p max +p min
[0027] Among them, p max is the maximum value of load data, p min is the minimum load value.
[0028] Based on the same inventive concept, the present invention also provides a regional short-term load forecasting system based on historical load characteristic optimization, comprising:
[0029] An initialization module is used to collect the load data of the area to be predicted every hour, perform Min-Max normalization on the load data, and record the maximum and minimum values of the load;
[0030] A matrix construction module is used to obtain load characteristics and construct a load characteristic matrix based on the preprocessed load data;
[0031] A model prediction module is used to train a CNN-LSTM combined model using the load feature matrix, input the load features after CNN processing into the LSTM to extract time series features, and perform prediction;
[0032] The final prediction module is used to restore the prediction results of the CNN-LSTM combination model based on the maximum and minimum values of the load data to obtain the final prediction value.
[0033] Furthermore, the matrix building module includes:
[0034] Obtain the hourly regional load sequence K+1 years before the forecast date, and calculate the daily average load sequence K years before the forecast date in units of days, with a sequence length of 365(K+1);
[0035] Select the historical daily average load K days before the forecast date to form a row vector with a length of K, which is used as the recent load characteristic row vector, denoted as c1;
[0036] Select the daily average load of the day before the forecast date and the same day of the previous K-1 weeks as the weekly load characteristic row vector, denoted as c2;
[0037] Select the daily average load of the day before the forecast date and the same day of the previous K-1 months as the monthly load characteristic row vector, denoted as c3;
[0038] Select the daily average load on the day before the forecast date and the same day in the previous K-1 years as the annual load characteristic row vector, denoted as c4;
[0039] Select the historical daily average load sequence of K+1 weeks before the day to be predicted, and calculate the weekly average load on a weekly basis to obtain a load sequence with a length of K+1. Select the first K values and subtract the last value to form a difference sequence with a length of K. The obtained difference sequence is normalized by Min-Max and used as the weekly load change feature, recorded as c5, to characterize the impact of human activities on the load on a weekly scale;
[0040] Select the historical daily average load sequence of K+1 months before the forecast date, and calculate the monthly average load on a monthly basis to obtain a load sequence with a length of K+1. Select the first K values and subtract the last value to form a difference sequence with a length of K. The obtained difference sequence is normalized by Min-Max and used as the monthly load change feature, recorded as c6, to characterize the impact of human activities on the load on a monthly scale;
[0041] Select the historical daily average load sequence of K+1 years before the forecast date, and calculate the annual average load on an annual basis to obtain a load sequence with a length of K+1. Select the first K values and subtract the last value to form a difference sequence with a length of K. The obtained difference sequence is normalized by Min-Max and used as the annual load change feature, recorded as c7, to characterize the impact of human activities on the load on an annual scale;
[0042] Stack the eigenvectors c1 to c7 row by row to get the load characteristic matrix Its dimension is 7×K.
[0043] Furthermore, the model prediction module includes:
[0044] Construct a CNN-LSTM model. CNN uses a one-dimensional convolutional neural network. CNN and LSTM are cascaded, and the last time step of the LSTM network is taken as the prediction result.
[0045] Train the CNN-LSTM model, the model input is the load feature matrix, the model output is the predicted load, and the training loss is the MSE error.
[0046] Furthermore, the final prediction module includes:
[0047] Suppose the prediction result of the CNN-LSTM combination model is x ′ pre , then the final predicted value x pre for:
[0048] x pre =x ′ pre p max +p min
[0049] Among them, pmax is the maximum value of load data, p min is the minimum load value.
[0050] Based on the same inventive concept, the present invention also provides a computing device, comprising: one or more processors, one or more memories and one or more programs, wherein the programs are stored in the memories and configured to be executed by the processors, and when the programs are loaded into the processors, the steps of the regional short-term load forecasting method based on historical load characteristic optimization as described in any one of the above items are implemented.
[0051] Based on the same inventive concept, the present invention also provides a storage medium, which stores a computer program, and the computer program includes program instructions. When the program instructions are executed by a processor, the processor executes the steps of a regional short-term load forecasting method based on historical load characteristic optimization for calculating the limit cut-off angle of a grid-type converter system according to any of the above items.
[0052] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. Most of the current regional short-term prediction methods use weather data, daily type data, and historical load data as features. However, there is a large uncertainty in the correlation between short-term load and weather and daily type data. Simply sending weather and daily type data into the prediction network often cannot improve the prediction effect, and may even increase negative effects. The current method requires complex data feature engineering from weather and daily type data; the present invention only requires a single historical load data to complete short-term load prediction; 2. The present invention performs a large amount of feature optimization on historical load data, and mines load features from different time scales, such as weekly load difference sequence characterizing the load change trend on the weekly scale, monthly load difference sequence characterizing the load change trend on the monthly scale, and annual load difference sequence characterizing the load change trend on the annual scale. These change trends can be understood from a certain perspective as the impact of meteorological, economic, human activities and other dimensions on load on different time scales. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 The figure is a flow chart of a method according to an embodiment of the present invention. DETAILED DESCRIPTION
[0054] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.
[0055] As attached Figure 1 As shown, the regional short-term load forecasting method based on historical load characteristic optimization of this embodiment includes:
[0056] S1: Collect the load data of the area to be predicted every hour, perform Min-Max normalization on the load data, and record the maximum and minimum values of the load;
[0057] S2: Based on the preprocessed load data, obtain load characteristics and construct a load characteristic matrix;
[0058] S3: Use the load feature matrix to train the CNN-LSTM combined model, input the load features after CNN processing into the LSTM to extract time series features and make predictions;
[0059] S4: Based on the maximum and minimum values of the load data, the prediction results of the CNN-LSTM combined model are restored to obtain the final prediction value.
[0060] Specifically, step S1, normalize the load data of the area every hour using the maximum-minimum method, and record the maximum value of the load as p max , the minimum load value is p min ;
[0061] In step S2, the load characteristic matrix is constructed:
[0062] S2.1. Obtain the hourly regional load sequence K+1 years before the date to be predicted, and calculate the daily average load sequence K years before the date to be predicted in days, with a sequence length of 365 (K+1); the daily average load here is obtained by calculating the average of all hourly loads every day, and the sequence length is 365 days (assuming it is not a leap year); all calculations in subsequent steps are based on the daily average load sequence extracted from the original load data;
[0063] S2.2, select the historical daily average load K days before the forecast date to form a row vector with a length of K, which is used as the recent load characteristic row vector, denoted as c1;
[0064] S2.3, select the daily average load on the day before the forecast date and the same day of the previous K-1 weeks as the weekly load characteristic row vector, denoted as c2;
[0065] S2.4, select the daily average load on the day before the forecast date and the same day of the previous K-1 months as the monthly load characteristic row vector, denoted as c3;
[0066] S2.5, select the daily average load on the day before the forecast date and the same day in the previous K-1 years as the annual load characteristic row vector, denoted as c4;
[0067] S2.6. Select the historical daily average load sequence of K+1 weeks before the day to be predicted, and calculate the weekly average load on a weekly basis to obtain a load sequence of length K+1. Select the first K values and subtract the last value (i.e., the average load value of the week before the day to be predicted) to form a difference sequence of length K. Perform maximum-minimum normalization on the obtained difference sequence as the weekly load change feature, recorded as c5, to characterize the impact of human activities on the load on a weekly scale;
[0068] S2.7. Select the historical daily average load sequence of K+1 months before the day to be predicted, and calculate the monthly average load on a monthly basis to obtain a load sequence of length K+1. Select the first K values and subtract the last value to form a difference sequence of length K. Perform maximum-minimum normalization on the obtained difference sequence as the monthly load change feature, recorded as c6, to characterize the impact of human activities on the load on a monthly scale;
[0069] S2.8. Select the historical daily average load sequence of K+1 years before the day to be predicted, and calculate the annual average load on an annual basis to obtain a load sequence of length K+1. Select the first K values and subtract the last value to form a difference sequence of length K. Perform maximum-minimum normalization on the obtained difference sequence as the annual load change feature, recorded as c7, to characterize the impact of human activities on the load on an annual scale;
[0070] S2.9. Stack the above characteristics by row to obtain the load characteristic matrix Its dimension is 7×K;
[0071] In this embodiment, the K value is generally selected between 3 and 5.
[0072] In step S3, the load feature matrix is used to train the CNN-LSTM combined model; including:
[0073] S3.1. Construct a CNN-LSTM model. CNN uses a one-dimensional convolutional neural network. CNN and LSTM are in cascade form. The last time step of the LSTM network is taken as the prediction result.
[0074] S3.2, train the CNN-LSTM model, the model input is the load feature matrix, the model output is the predicted load, and the training loss is the MSE error;
[0075] In step S4, the prediction data of the CNN-LSTM combined model is restored according to the maximum and minimum loads recorded in S1. When the output of CNN-LSTM is x ′ pre When the final prediction value is x pre =x ′ pre p max +pmin ;
[0076] Based on the same inventive concept, this embodiment also provides a regional short-term load forecasting system based on historical load characteristic optimization, including:
[0077] An initialization module is used to collect the load data of the area to be predicted every hour, perform Min-Max normalization on the load data, and record the maximum and minimum values of the load;
[0078] A matrix construction module is used to obtain load characteristics and construct a load characteristic matrix based on the preprocessed load data;
[0079] A model prediction module is used to train a CNN-LSTM combined model using the load feature matrix, input the load features after CNN processing into the LSTM to extract time series features, and perform prediction;
[0080] The final prediction module is used to restore the prediction results of the CNN-LSTM combination model based on the maximum and minimum values of the load data to obtain the final prediction value.
[0081] Furthermore, the matrix building module includes:
[0082] Obtain the hourly regional load sequence K+1 years before the forecast date, and calculate the daily average load sequence K years before the forecast date in units of days, with a sequence length of 365(K+1);
[0083] Select the historical daily average load K days before the forecast date to form a row vector with a length of K, which is used as the recent load characteristic row vector, denoted as c1;
[0084] Select the daily average load of the day before the forecast date and the same day of the previous K-1 weeks as the weekly load characteristic row vector, denoted as c2;
[0085] Select the daily average load of the day before the forecast date and the same day of the previous K-1 months as the monthly load characteristic row vector, denoted as c3;
[0086] Select the daily average load on the day before the forecast date and the same day in the previous K-1 years as the annual load characteristic row vector, denoted as c4;
[0087] Select the historical daily average load sequence of K+1 weeks before the day to be predicted, and calculate the weekly average load on a weekly basis to obtain a load sequence with a length of K+1. Select the first K values and subtract the last value to form a difference sequence with a length of K. The obtained difference sequence is normalized by Min-Max and used as the weekly load change feature, recorded as c5, to characterize the impact of human activities on the load on a weekly scale;
[0088] Select the historical daily average load sequence of K+1 months before the forecast date, and calculate the monthly average load on a monthly basis to obtain a load sequence with a length of K+1. Select the first K values and subtract the last value to form a difference sequence with a length of K. The obtained difference sequence is normalized by Min-Max and used as the monthly load change feature, recorded as c6, to characterize the impact of human activities on the load on a monthly scale;
[0089] Select the historical daily average load sequence of K+1 years before the forecast date, and calculate the annual average load on an annual basis to obtain a load sequence with a length of K+1. Select the first K values and subtract the last value to form a difference sequence with a length of K. The obtained difference sequence is normalized by Min-Max and used as the annual load change feature, recorded as c7, to characterize the impact of human activities on the load on an annual scale;
[0090] Stack the eigenvectors c1 to c7 row by row to get the load characteristic matrix Its dimension is 7×K.
[0091] Furthermore, the model prediction module includes:
[0092] Construct a CNN-LSTM model. CNN uses a one-dimensional convolutional neural network. CNN and LSTM are cascaded, and the last time step of the LSTM network is taken as the prediction result.
[0093] Train the CNN-LSTM model, the model input is the load feature matrix, the model output is the predicted load, and the training loss is the MSE error.
[0094] Furthermore, the final prediction module includes:
[0095] Suppose the prediction result of the CNN-LSTM combination model is x ′ pre , then the final predicted value x pre for:
[0096] x pre =x ′ pre p max +p min
[0097] Among them, p max is the maximum value of load data, p min is the minimum load value.
[0098] Based on the same inventive concept, this embodiment also provides a computing device, including: one or more processors, one or more memories and one or more programs, wherein the programs are stored in the memories and configured to be executed by the processors, and when the programs are loaded into the processors, the steps of the regional short-term load forecasting method based on historical load characteristic optimization according to any one of the above items are implemented.
[0099] Based on the same inventive concept, this embodiment also provides a storage medium, which stores a computer program, and the computer program includes program instructions. When the program instructions are executed by a processor, the processor executes the steps of the regional short-term load forecasting method based on historical load characteristic optimization for calculating the limit cut-off angle of the grid-type inverter system according to any of the above items.
[0100] The present invention constructs and optimizes relevant features for historical loads that are highly correlated with power system loads, thereby improving the effect of short-term load forecasting for the power system.
[0101] The present invention provides a regional short-term load forecasting method and system based on historical load feature optimization, which ensures the accuracy of data preprocessing through detailed Min-Max normalization processing and recording the maximum and minimum values of the load; constructs a multi-time scale load feature matrix to comprehensively capture the load change trend; adopts a CNN-LSTM combined model to improve the prediction accuracy and robustness of the model; uses the maximum and minimum values of data preprocessing to accurately restore the prediction results; and improves the model's adaptability to sudden events through dynamic updating of multi-time scale features and continuous training of the model. The solution to these problems significantly improves the accuracy and reliability of short-term power load forecasting, and provides strong support for the stable operation and reasonable scheduling of the power system.
Claims
1. A regional short-term load forecasting method based on historical load characteristic optimization, characterized in that: include: S1: Collect the load data of the area to be predicted every hour, perform Min-Max normalization on the load data, and record the maximum and minimum values of the load; S2: Based on the preprocessed load data, obtain load characteristics and construct a load characteristic matrix; S3: Use the load feature matrix to train the CNN-LSTM combined model, input the load features after CNN processing into the LSTM to extract time series features and make predictions; S4: Based on the maximum and minimum values of the load data, the prediction results of the CNN-LSTM combined model are restored to obtain the final prediction value.
2. The regional short-term load forecasting method based on historical load characteristic optimization according to claim 1 is characterized in that: The step S2 comprises: S2.1: Obtain the hourly regional load sequence K+1 years before the forecast date, and calculate the daily average load sequence K years before the forecast date in units of days, with a sequence length of 365(K+1); S2.2: Select the historical daily average load K days before the day to be predicted to form a row vector with a length of K, which is used as the recent load characteristic row vector, denoted as c1; S2.3: Select the daily average load on the day before the forecast date and the same day in the previous K-1 weeks as the weekly load characteristic row vector, denoted as c2; S2.4: Select the daily average load on the day before the forecast date and the same day of the previous K-1 months as the monthly load characteristic row vector, denoted as c3; S2.5: Select the daily average load on the day before the forecast date and the same day in the previous K-1 years as the annual load characteristic row vector, denoted as c4; S2.6: Select the historical daily average load sequence of K+1 weeks before the day to be predicted, and calculate the weekly average load on a weekly basis to obtain a load sequence of length K+1. Select the first K values and subtract the last value to form a difference sequence of length K. Perform Min-Max normalization on the obtained difference sequence as the weekly load change feature, recorded as c5, to characterize the impact of human activities on the load on a weekly scale; S2.7: Select the historical daily average load sequence of K+1 months before the forecast date, and calculate the monthly average load on a monthly basis to obtain a load sequence of length K+1. Select the first K values and subtract the last value to form a difference sequence of length K. Perform Min-Max normalization on the obtained difference sequence as the monthly load change feature, recorded as c6, to characterize the impact of human activities on the load on a monthly scale; S2.8: Select the historical daily average load sequence of K+1 years before the day to be predicted, and calculate the annual average load on an annual basis to obtain a load sequence of length K+1. Select the first K values and subtract the last value to form a difference sequence of length K. Perform Min-Max normalization on the obtained difference sequence as the annual load change feature, recorded as c7, to characterize the impact of human activities on the load on an annual scale; S2.9: Stack the features calculated in steps S2.2 to S2.8 by row to obtain a load feature matrix Its dimension is 7×K.
3. The regional short-term load forecasting method based on historical load characteristic optimization according to claim 1 is characterized in that: The step S3 comprises: S3.1: Construct a CNN-LSTM model. CNN uses a one-dimensional convolutional neural network. CNN and LSTM are cascaded, and the last time step of the LSTM network is taken as the prediction result. S3.2: Train the CNN-LSTM model. The model input is the load feature matrix, the model output is the predicted load, and the training loss is the MSE error.
4. The regional short-term load forecasting method based on historical load characteristic optimization according to claim 1 is characterized in that: The step S4 comprises: Suppose the prediction result of the CNN-LSTM combination model is x ′ pre , then the final predicted value x pre for: x pre =x ′ pre p max +p min Among them, p max is the maximum value of load data, p min is the minimum load value.
5. A regional short-term load forecasting system based on historical load characteristic optimization, characterized in that: include: An initialization module is used to collect the load data of the area to be predicted every hour, perform Min-Max normalization on the load data, and record the maximum and minimum values of the load; A matrix construction module is used to obtain load characteristics and construct a load characteristic matrix based on the preprocessed load data; A model prediction module is used to train a CNN-LSTM combined model using the load feature matrix, input the load features after CNN processing into the LSTM to extract time series features, and perform prediction; The final prediction module is used to restore the prediction results of the CNN-LSTM combination model based on the maximum and minimum values of the load data to obtain the final prediction value.
6. The regional short-term load forecasting system based on historical load characteristic optimization according to claim 5 is characterized in that: The matrix building module includes: Obtain the hourly regional load sequence K+1 years before the forecast date, and calculate the daily average load sequence K years before the forecast date in units of days, with a sequence length of 365(K+1); Select the historical daily average load K days before the forecast date to form a row vector with a length of K, which is used as the recent load characteristic row vector, denoted as c1; Select the daily average load of the day before the forecast date and the same day of the previous K-1 weeks as the weekly load characteristic row vector, denoted as c2; Select the daily average load of the day before the forecast date and the same day of the previous K-1 months as the monthly load characteristic row vector, denoted as c3; Select the daily average load on the day before the forecast date and the same day in the previous K-1 years as the annual load characteristic row vector, denoted as c4; Select the historical daily average load sequence of K+1 weeks before the day to be predicted, and calculate the weekly average load on a weekly basis to obtain a load sequence with a length of K+1. Select the first K values and subtract the last value to form a difference sequence with a length of K. The obtained difference sequence is normalized by Min-Max and used as the weekly load change feature, recorded as c5, to characterize the impact of human activities on the load on a weekly scale; Select the historical daily average load sequence of K+1 months before the forecast date, and calculate the monthly average load on a monthly basis to obtain a load sequence with a length of K+1. Select the first K values and subtract the last value to form a difference sequence with a length of K. The obtained difference sequence is normalized by Min-Max and used as the monthly load change feature, recorded as c6, to characterize the impact of human activities on the load on a monthly scale; Select the historical daily average load sequence of K+1 years before the forecast date, and calculate the annual average load on an annual basis to obtain a load sequence with a length of K+1. Select the first K values and subtract the last value to form a difference sequence with a length of K. The obtained difference sequence is normalized by Min-Max and used as the annual load change feature, recorded as c7, to characterize the impact of human activities on the load on an annual scale; Stack the eigenvectors c1 to c7 row by row to get the load characteristic matrix Its dimension is 7×K.
7. The regional short-term load forecasting system based on historical load characteristic optimization according to claim 6 is characterized in that: The model prediction module includes: Construct a CNN-LSTM model. CNN uses a one-dimensional convolutional neural network. CNN and LSTM are cascaded, and the last time step of the LSTM network is taken as the prediction result. Train the CNN-LSTM model, the model input is the load feature matrix, the model output is the predicted load, and the training loss is the MSE error.
8. The regional short-term load forecasting system based on historical load characteristic optimization according to claim 6 is characterized in that: The final prediction module includes: Suppose the prediction result of the CNN-LSTM combination model is x ′ pre , then the final predicted value x pre for: x pre =x ′ pre p max +p min Among them, p max is the maximum value of load data, p min is the minimum load value.
9. A computing device, characterized in that include: One or more processors, one or more memories and one or more programs, wherein the programs are stored in the memories and are configured to be executed by the processors, and when the programs are loaded into the processors, the steps of the regional short-term load forecasting method based on historical load characteristic optimization according to any one of claims 1 to 4 are implemented.
10. A storage medium, characterized in that: The storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, the processor executes the steps of the regional short-term load forecasting method based on historical load characteristic optimization according to any one of claims 1 to 4.