A method for regional power load forecasting

Abstract

The invention relates to the technical field of smart grids and discloses a method for regional power load forecasting, comprising collecting historical power consumption data; constructing a KANs model for power load forecasting, using a part of time characteristics, meteorological characteristics and power consumption data in the historical data as a training set and another part as a test set; inputting the training set into the KANs model to obtain an initial power load forecasting model; inputting the test set into the initial power load forecasting model for verification to obtain a trained power load forecasting model, inputting the time characteristics and meteorological characteristics of a to-be-forecasted area in a specific time period into the power load forecasting model to obtain a power load forecasting result; the method can be applied to the forecasting of power loads at various time granularities and has the characteristic of high accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grids, and in particular to a method for forecasting regional power load. Background Art

[0002] In recent years, with the widespread adoption of power trading and the promotion of medium- and long-term, as well as spot, power trading, electricity retailers and power trading centers have emerged on a large scale. Accurately predicting the power consumption of various types of users in a region, as well as future trends, is crucial for power retailers, power grids, and power generation companies involved in power trading. Therefore, employing an effective power sales system to accurately predict regional user load plays a crucial role.

[0003] With the development of photovoltaic, wind power and energy storage technologies, traditional power grids are changing with each passing day. New concepts such as virtual power plants, smart grids and microgrids are constantly emerging, which pose challenges to traditional power load forecasting. Taking into account meteorological factors such as temperature, humidity, wind speed, precipitation, cloud cover, lighting conditions and extreme weather, which have a significant impact on user power load, the traditional method uses similar day method and physical method to predict power load.

[0004] The similar day method is a forecasting method based on historical data. Its core concept is to predict future power load by selecting historical days similar to the day being forecasted (similar days). While this method considers the relationship between power load and meteorological factors (such as temperature and humidity), as well as power load trends over different time periods, measuring the similarity of various characteristics between two dates is a major challenge, often requiring extensive computation and time-consuming statistical analysis. Furthermore, some extreme weather conditions are difficult to find historically comparable dates.

[0005] The physical method uses linear or nonlinear regression to determine the correlation factors of the impact of different characteristic factors on power load by controlling for variables. However, with high-dimensional features, controlling for variables results in sparse data points, making it difficult to perform regression on the coefficients of the correlation factors. Furthermore, the physical method faces the challenge of sparse data during extreme weather events, making it difficult to quantify the impact of extreme characteristics on user load. This results in low prediction accuracy.

[0006] The current electricity market is divided into medium- and long-term markets (specifically, monthly and annual) and the day-ahead spot market in terms of time. In terms of space, users are grouped together into a single entity, based on provinces (autonomous regions or municipalities). In this context, an effective method is needed to collectively forecast regional electricity user load at multiple time granularities. Summary of the Invention

[0007] The present invention aims to provide a regional power load forecasting method to meet the demand for collective forecasting of regional power user loads at multiple time granularities.

[0008] To achieve the above object, the present invention adopts a regional power load forecasting method, comprising the following steps: Step 1: Collect historical electricity consumption data of regional users with hourly statistical accuracy; collect the time characteristics corresponding to the electricity consumption and the meteorological characteristics of the time period corresponding to the electricity consumption; Step 2: Construct the KANs model for power load forecasting through the Kolmogorov-Arnold representation theorem. Step 3: Use part of the time characteristics, meteorological characteristics, and electricity consumption data in step 1 as a training set, and the other part as a test set; Step 4: Input the training set into the KANs model for training to obtain the initial power load forecasting model; Step 5: Input the test set into the initial power load forecasting model for verification, optimize the parameters of the initial power load forecasting model by cross-validation and setting a loss function, and obtain a trained power load forecasting model; Step 6: Input the time characteristics and meteorological characteristics of the area to be predicted in a specific time period into the trained power load forecasting model to obtain the power load forecast result of the area in the specific time period.

[0009] Furthermore, the time characteristics of step 1 include the hour period, date, week and month corresponding to the power consumption.

[0010] Furthermore, the meteorological characteristics in step 1 include temperature, humidity, wind speed, precipitation, cloud cover and light conditions, and correlation analysis is performed between the meteorological characteristics and electricity consumption to obtain meteorological characteristics that have a causal relationship with electricity consumption; Furthermore, between step 1 and step 2, the temporal and meteorological characteristics of the historical data are preprocessed. The specific preprocessing method is: the missing values ​​in the historical data are filled with the mean, and the mean-standard deviation normalization is performed to make the temporal and meteorological characteristics conform to the input format of the KANs model.

[0011] Furthermore, the KANs model selected in step three includes an input layer, an MLP layer, an RKANs layer, an LSTM layer and an output layer; the input layer is used to receive multidimensional data of time features and meteorological features; the MLP layer is used for preliminary feature extraction and linear transformation; the RKANs layer is used to capture complex nonlinear relationships; the LSTM layer is used to process time series dependencies and extract long-term dependency features in time series; and the output layer is used to output prediction results.

[0012] Furthermore, in step 4, the KANs model includes a trainable activation function. The training process of the activation function includes the following steps: first, the activation function is initialized with a set of trainable random parameters; then, forward propagation is performed to generate a function by linearly transforming the input value x; then, the output value of the activation function is calculated; next, backpropagation is performed: the gradient of the loss with respect to the activation function parameters is calculated; finally, the parameters of the activation function are updated using the Adam or SGD optimization algorithm.

[0013] Furthermore, the cross-validation in step 5 is implemented using the cross-validation tool in sklearn.model_selection. Specifically, the training set and test set are divided into K subsets, and then K training and validation are performed. Each time, one subset is selected as the validation set, and the rest are used as the training set.

[0014] sklearn.model_selection is a module in Python's scikit-learn library that provides tools for model selection and evaluation.

[0015] Furthermore, the loss function in step 5 is selected as the mean square error function, and the expression of the mean square error function is: , where N is the number of data points, is the actual value of the power load at time t, is the predicted value of the power load at time t.

[0016] Finally, after step six, the method also includes a step of continuously optimizing the power load forecasting model, in which the performance of the power load forecasting model is evaluated by the mean square error or the mean absolute error; the structure and parameters of the power load forecasting model are optimized according to the evaluation results to improve the prediction performance of the KANs prediction model.

[0017] The principles and advantages of this method are as follows: It uses the KANs model as a forecasting tool. The KANs model can take multidimensional data that influence power load results as input parameters, achieving high accuracy in data fitting and partial differential solutions. This makes forecasts more objective and accurate. Furthermore, it is applicable to power load forecasting for smart grids and microgrids comprised of photovoltaic, wind power, and energy storage, as well as for virtual grids.

[0018] This method incorporates multiple meteorological factors such as temperature, humidity, and wind speed into the prediction model, which can improve the KANs model's responsiveness to user electricity consumption behavior and overcome the problem in the traditional similar day method that some extreme weather conditions cannot be used to predict power load status because of the inability to find corresponding regular power load reference data.

[0019] This method uses the inherent causal relationship between time characteristics and meteorological characteristics to dynamically adjust the weight of each meteorological impact, thereby effectively outputting a forecast of power load due to extreme weather.

[0020] Due to the trainable activation function in the KANs model in this method, the KAN model has the ability to generalize to arbitrary width and depth. It can be used in power forecasting with both time factors and meteorological factors, and can meet the needs of collective forecasting of regional power user loads at multiple time granularities. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of the method for regional power load forecasting in Example 1 of the present invention. DETAILED DESCRIPTION

[0022] The following is further described in detail through specific implementation methods: Example 1 Basically as attached Figure 1 As shown, a method for regional power load forecasting includes the following steps: Step 1 S1: Collect historical electricity consumption data of regional users with statistical accuracy in hours; collect time characteristics corresponding to electricity consumption and meteorological characteristics of the time period corresponding to electricity consumption; Step 2 S2: Construct the KANs model for power load forecasting through the Kolmogorov-Arnold representation theorem. Step 3 S3: Use part of the time characteristics, meteorological characteristics and electricity consumption data in step 1 as a training set and the other part as a test set; Step 4 S4: input the training set into the KANs model for training to obtain the initial power load forecasting model; Step 5 S5, inputting the test set into the initial power load forecasting model for verification, optimizing the parameters in the initial power load forecasting model by cross-validation and setting a loss function, and obtaining a trained power load forecasting model; Step 6 S6: Input the time characteristics and meteorological characteristics of the area to be predicted in the specific time period into the trained power load prediction model to obtain the power load prediction result of the area in the specific time period.

[0023] The time characteristics of step 1 S1 include the hour period, date, week and month corresponding to the power consumption.

[0024] The meteorological characteristics of step 1 S1 include temperature, humidity, wind speed, precipitation, cloud cover and light conditions, and the correlation analysis between meteorological characteristics and electricity consumption is performed to obtain meteorological characteristics that have a causal relationship with electricity consumption; Between step 1 S1 and step 2 S2, the temporal and meteorological characteristics in the historical data are also preprocessed. The specific preprocessing method is: the missing values ​​in the historical data are filled with the mean, and the mean-standard deviation normalization is performed to make the temporal and meteorological characteristics conform to the input format of the KANs model.

[0025] The KANs model selected in step 3 (S3) consists of an input layer, an MLP layer, an RKANs layer, an LSTM layer, and an output layer. The input layer receives multidimensional data on temporal and meteorological characteristics; the MLP layer performs preliminary feature extraction and linear transformation; the RKANs layer captures complex nonlinear relationships; the LSTM layer processes temporal dependencies and extracts long-term dependency features from the time series; and the output layer outputs the prediction results.

[0026] In step 4 S4, the KANs model includes a trainable activation function. The activation function training process includes the following steps: first, the activation function is initialized with a set of trainable random parameters; then, forward propagation is performed to generate a function by linearly transforming the input value x; then, the output value of the activation function is calculated; next, backpropagation is performed: the gradient of the loss with respect to the activation function parameters is calculated; finally, the parameters of the activation function are updated using the Adam or SGD optimization algorithm.

[0027] The cross-validation in step 5 S5 is implemented using the cross-validation tool in sklearn.model_selection. Specifically, the training set and test set are divided into K subsets, and then K training and validation are performed. Each time, one subset is selected as the validation set, and the rest are used as the training set.

[0028] sklearn.model_selection is a module in Python's scikit-learn library that provides tools for model selection and evaluation.

[0029] The loss function in step 5 S5 is the mean square error function, and the expression of the mean square error function is: , where N is the number of data points, is the actual value of the power load at time t, is the predicted value of the power load at time t.

[0030] After step six S6, it also includes a step of continuously optimizing the power load forecasting model, which evaluates the performance of the power load forecasting model through mean square error or mean absolute error; based on the evaluation results, the structure and parameters of the power load forecasting model are optimized to improve the prediction performance of the KANs prediction model.

[0031] Example 2: This method is used in the power load forecasting of virtual power plants, smart grids and microgrids. The KANs prediction model in this method is compiled into a computer executable program to form a power load forecasting model that can run automatically.

[0032] Leveraging the data transmission and storage capabilities of computers, historical electricity data is collected and converted into electronic data. This electronic data is then pre-processed to serve as training data for the KANs prediction model. This pre-processing method fills missing values ​​in the historical electricity data with their mean and normalizes them using mean-standard deviation to ensure that their temporal and meteorological characteristics conform to the KANs model input format.

[0033] The KANs prediction model is trained using historical data, and then the time characteristics corresponding to the power load in a certain area that needs to be predicted, such as which hour, day of the week, and which month, as well as meteorological characteristics, such as temperature, humidity, wind speed, precipitation, cloud cover, and lighting conditions in a certain time period in the future, are used as input parameters of the power load prediction model. After the computer runs the program, the power load prediction model completes the power load prediction results that need to be predicted.

[0034] The above is only an embodiment of the present invention, and the common knowledge such as the specific technical solutions and / or characteristics in the solution are not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the description can be used to interpret the content of the claims.

Claims

1. A method for predicting regional power load, characterized in that: The following steps are included: Step 1: Collect historical electricity consumption data of regional users with statistical accuracy in hours; collect time characteristics corresponding to electricity consumption and meteorological characteristics of the time period corresponding to electricity consumption; Step 2: Construct the KANs model for power load forecasting through the Kolmogorov-Arnold representation theorem. Step 3: Use part of the time characteristics, meteorological characteristics, and electricity consumption data in step 1 as a training set, and the other part as a test set; Step 4: Input the training set into the KANs model for training to obtain the initial power load forecasting model; Step 5: Input the test set into the initial power load forecasting model for verification, optimize the parameters in the initial power load forecasting model by cross-validation and setting loss function, and obtain a trained power load forecasting model; Step 6: Input the time characteristics and meteorological characteristics of the area to be predicted in a specific time period into the trained power load forecasting model to obtain the power load forecast result of the area in the specific time period.

2. The method for regional power load forecasting according to claim 1, characterized in that: The time characteristics of step one include the hour period, date, week and month corresponding to the power consumption.

3. The method for regional power load forecasting according to claim 2, characterized in that: The meteorological characteristics in step one include temperature, humidity, wind speed, precipitation, cloud cover and light conditions, and a correlation analysis is performed between the meteorological characteristics and electricity consumption to obtain meteorological characteristics that have a causal relationship with electricity consumption.

4. The method for regional power load forecasting according to claim 3, characterized in that: Between step one and step two, the time characteristics and meteorological characteristics in the historical data are also preprocessed. The specific preprocessing method is: the missing values ​​in the historical data are filled with the mean, and the mean-standard deviation normalization is performed to make the time characteristics and meteorological characteristics conform to the input format of the KANs model.

5. The method for regional power load forecasting according to claim 4, characterized in that: The KANs model selected in step three includes an input layer, an MLP layer, an RKANs layer, an LSTM layer and an output layer; the input layer is used to receive multidimensional data of time characteristics and meteorological characteristics; the MLP layer is used for preliminary feature extraction and linear transformation; the RKANs layer is used to capture complex nonlinear relationships; the LSTM layer is used to process time series dependencies and extract long-term dependency features in time series, and the output layer is used to output prediction results.

6. The method for regional power load forecasting according to claim 5, characterized in that: The cross validation in step 5 is implemented using the cross validation tool in sklearn.model_selection. Specifically, the training set and the test set are divided into K subsets, and then K training and validation are performed, each time selecting one subset as the validation set and the rest as the training set.

7. The method for regional power load forecasting according to claim 6, characterized in that: The loss function in step 5 is the mean square error function, and the expression of the mean square error function is: , where N is the number of data points, is the actual value of the power load at time t, is the predicted value of the power load at time t.

8. The method for regional power load forecasting according to claim 7, characterized in that: After step six, the method also includes a step of continuously optimizing the power load forecasting model, in which the performance of the power load forecasting model is evaluated by mean square error or mean absolute error; the structure and parameters of the power load forecasting model are optimized according to the evaluation results to improve the prediction performance of the KANs prediction model.

Images