Short-term power load prediction model construction method and prediction method

Abstract

The invention is suitable for the field of short-term power load prediction model construction methods, and discloses a short-term power load prediction model construction method and a short-term power load prediction method. The historical data of the power system comprises a plurality of historical load data of the power system and corresponding historical meteorological environment data; clustering the plurality of historical meteorological environment data to obtain a plurality of meteorological clusters; inputting the historical meteorological environment data of the plurality of meteorological clusters into a time convolutional network (TCN) to extract meteorological environment characteristics of the meteorological clusters, and generating a plurality of training samples according to the extracted meteorological environment characteristics and the corresponding historical load data of the power system; a plurality of training samples are input into an improved Autoformer model for training, a short-term power system load prediction model is obtained, and the short-term power load prediction model constructed through the method is more accurate in short-term power load prediction.

Description

Technical Field

[0001] The present invention relates to the field of power load forecasting, and in particular to a short-term power load forecasting model construction method and a forecasting method. Background Art

[0002] With the development of smart grid technology, the system needs to be able to quickly respond to changes in power demand. Power load forecasting can make the power grid more flexible and efficient in power distribution and management. Moreover, power load forecasting is the basis of demand response and energy management. By forecasting load, demand-side management can be better performed and the allocation and use of power resources can be optimized. Accurate power load forecasting is crucial to the safe, stable and economic operation of the power system. It helps to avoid insufficient or excessive power supply, reduce energy waste and improve the economic benefits of the power grid.

[0003] Meteorological factors such as temperature, humidity, wind speed and precipitation have a significant impact on power load, but the uncertainty and volatility of these factors increase the difficulty of prediction. Traditional load forecasting methods can no longer meet the requirements of modern power systems for prediction accuracy and efficiency.

[0004] Therefore, the existing technology still needs to be improved and developed. Summary of the invention

[0005] The first object of the present invention is to provide a method for constructing a short-term power load forecasting model, which constructs a short-term power load forecast based on an improved Autoformer model and can improve the prediction accuracy of short-term power system load.

[0006] To achieve the above object, the solution provided by the present invention is:

[0007] A method for constructing a short-term power load forecasting model, the method comprising: collecting historical data of a power system, the historical data of the power system comprising a plurality of historical load data of the power system and corresponding historical meteorological environment data; clustering the plurality of historical meteorological environment data to obtain a plurality of meteorological clusters; inputting the historical meteorological environment data of the plurality of meteorological clusters into a temporal convolutional network (TCN) to extract meteorological environment features of the meteorological clusters, and generating a plurality of training samples according to the extracted meteorological environment features and the corresponding historical load data of the power system; and inputting the plurality of training samples into an improved Autoformer model for training to obtain a short-term power system load forecasting model.

[0008] Preferably, when clustering the plurality of historical meteorological environment data to obtain the plurality of meteorological clusters, the fuzzy C-means clustering algorithm is used to cluster the plurality of historical meteorological environment data to obtain the plurality of meteorological clusters. The fuzzy C-means clustering algorithm is used to cluster the plurality of historical meteorological environment data to obtain the plurality of meteorological clusters, including: obtaining the number of clusters C, and randomly initializing C cluster centers; for each historical meteorological environment data, calculating its membership to each cluster center; according to the plurality of historical meteorological environment data||x i -c k || The membership of the environmental data is calculated and each cluster center is updated; whether the updated cluster center converges is determined, and if so, the final cluster center and the membership matrix corresponding to the multiple historical meteorological environmental data are output; according to the membership matrix of the multiple historical meteorological environmental data, the cluster to which each historical meteorological environmental data belongs is determined according to the maximum membership principle in the fuzzy set, and multiple meteorological clusters are obtained.

[0009] Preferably, the temporal convolutional network TCN is composed of a plurality of stacked convolutional layers, each of which includes a causal convolution module, a dilated convolution module and a residual connection module. The causal convolution module is used to ensure the causality of the model when performing time series prediction, the dilated convolution module is used to capture longer-distance dependencies, and the residual connection module alleviates the gradient vanishing problem in deep network training by directly adding the output of the previous layer to the output of the current layer.

[0010] Preferably, the improved Autoformer model includes an encoder module, a decoder module and an output layer, the decoder module receives the decomposition components output by the encoder module, and decodes and reconstructs the decomposition components, the output layer receives the output results of the decoder module, and outputs the prediction results, the encoder module includes multiple MOEDecomp sequence decomposition layers and autocorrelation mechanism layers, and the decoder module includes a cumulative structure for processing trend periodic elements and a stacked autocorrelation structure for processing seasonal elements.

[0011] Preferably, the step of inputting the plurality of training samples into an improved Autoformer model for training to obtain a short-term power system load forecasting model comprises:

[0012] Inputting a plurality of training samples as long sequences into the encoder module, and obtaining a trend sequence and a seasonal feature sequence generated by the encoder module;

[0013] Inputting the generated trend sequence and seasonal feature sequence into the decoder module to obtain the output result of the decoder module;

[0014] Inputting the output result of the decoder module into the output layer to obtain the prediction result output by the output layer;

[0015] The loss function is calculated based on the prediction result output by the output layer and the actual load data, and the parameters of the improved Autoformer model are gradient calculated and updated according to the loss function through the back propagation algorithm to obtain a short-term power system load forecasting model.

[0016] Preferably, the MOEDecomp sequence decomposition layer is used to decompose the input training samples into trend components and seasonal components. The calculation formulas of the trend component and the seasonal component are respectively expressed as:

[0017] X trend =Soft max(L(x))*(F(x))

[0018] X seasonal =xX trend

[0019] Where Xtrend represents the trend component, Xseaonal represents the seasonal component, x represents the input training sample, F(x) is a set of average pooling filters, L(x) is the weight used to mix the trend, and the Softmax function is used to ensure that the sum of the weights is 1.

[0020] Preferably, the autocorrelation mechanism layer includes a period-based dependency discovery mechanism and a delay information aggregation mechanism. The period-based dependency discovery mechanism is used to identify similar subsequences with dependencies in a time series, and the delay information aggregation mechanism is used to align and aggregate information of similar subsequences.

[0021] The second object of the present invention is to provide a short-term power load forecasting method, which comprises: obtaining predicted meteorological environment data within the prediction time, and extracting features of the predicted meteorological environment data; inputting the extracted features into a short-term power system load forecasting model to obtain short-term power system load forecasting data, and the short-term power system load forecasting model adopts the short-term power system load forecasting model as described above.

[0022] Preferably, before extracting the features of the predicted meteorological environment data, the method further includes: preprocessing the acquired predicted meteorological environment data.

[0023] In the technical solution provided by the present invention, by clustering multiple historical meteorological and environmental data, the inherent structure and pattern in the data can be better discovered, which helps to improve the accuracy of clustering. Moreover, by extracting the feature vector of the meteorological and environmental data sample through the time convolution network TCN, it helps to capture the multi-level characteristics of the meteorological and environmental data. In addition, the improved Autoformer model uses a set of average filters of different sizes to capture multiple trend features in the input signal, and combines these trend features through data-related weights to form a final trend representation, which greatly effectively processes long time series with complex periodic patterns and trend components, helps to reduce the impact of the unstable characteristics of meteorological and environmental data on the prediction results, thereby improving the prediction accuracy of the short-term power system load forecasting model, and further improving the prediction accuracy of the short-term power system load. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0025] Figure 1 is a flow chart of a method for constructing a short-term power load forecasting model provided by an embodiment of the present invention;

[0026] Figure 2 It is a flow chart of a short-term power load forecasting method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0028] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 As shown, in an embodiment of the present invention, a method for constructing a short-term power load forecasting model includes:

[0029] S101, collecting historical data of the power system, where the historical data of the power system includes a plurality of historical load data of the power system and corresponding historical meteorological environment data;

[0030] S102, clustering a plurality of historical meteorological environment data to obtain a plurality of meteorological clusters;

[0031] S103, inputting the historical meteorological environment data of multiple meteorological clusters into the time convolution network TCN to extract the meteorological environment characteristics of the meteorological clusters, and generating multiple training samples according to the extracted meteorological environment characteristics and the corresponding historical load data of the power system;

[0032] S104, inputting multiple training samples into the improved Autoformer model for training to obtain a short-term power system load forecasting model.

[0033] In this embodiment, in step S101, the historical meteorological environment data includes temperature, humidity, total solar irradiance, direct solar irradiance, scattered solar irradiance, wind speed, wind direction, air pressure, microparticle index and other data.

[0034] The short-term power load forecasting model construction method of this embodiment, by clustering multiple historical meteorological and environmental data, can better discover the intrinsic structure and pattern in the data, which helps to improve the accuracy of clustering. Moreover, the feature vector of the meteorological and environmental data sample is extracted through the time convolution network TCN, which helps to capture the multi-level characteristics of the meteorological and environmental data. In addition, the improved Autoformer model uses a set of average filters of different sizes to capture multiple trend features in the input signal, and combines these trend features through data-related weights to form a final trend representation, which greatly effectively processes long time series with complex periodic patterns and trend components, helps to reduce the impact of the unstable characteristics of meteorological and environmental data on the forecast results, thereby improving the forecast accuracy of the short-term power system load forecasting model, and further improving the forecast accuracy of the short-term power system load.

[0035] In this embodiment, before clustering the historical meteorological environment data to obtain multiple meteorological clusters, the method further includes: preprocessing the historical meteorological environment data, specifically including data missing value processing, normalization processing and outlier processing.

[0036] In practical applications, there may be some missing values ​​in the power system historical data, such as data loss due to sensor failure or communication problems. The method of dealing with missing values ​​in data can be to delete the samples where the missing values ​​are located, or to use interpolation methods to fill the missing values.

[0037] There may be outliers in the historical data of the power system, that is, values ​​that are obviously inconsistent with other data or deviate from the norm. Outliers may be caused by sensor failure, human error or other abnormal conditions. The method of dealing with outliers is to delete the samples where the outliers are located, or to repair them using interpolation or smoothing techniques.

[0038] In this embodiment, in step S102, the historical meteorological environment data is clustered using a fuzzy C-means clustering algorithm to obtain a plurality of meteorological clusters.

[0039] Understandably, the fuzzy C-means clustering algorithm is a soft clustering algorithm based on fuzzy logic. It gives the degree to which the sample belongs to the class according to the similarity between the sample and the cluster center, so it can provide more flexible clustering results. The advantage of the fuzzy C-means clustering algorithm is that it can better handle the uncertainty and ambiguity of data and can discover overlapping clustering structures in the data.

[0040] In this embodiment, a fuzzy C-means clustering algorithm is used to cluster multiple historical meteorological environment data to obtain multiple meteorological clusters, which can better discover the inherent structure and pattern in the data and help improve the accuracy of clustering.

[0041] Specifically include:

[0042] S1021. Obtain the number of clusters C, and randomly initialize C cluster centers.

[0043] S1022. For each historical meteorological environment data, calculate its membership degree for each cluster center. The membership degree calculation formula is:

[0044]

[0045] In the formula, x i Represents historical meteorological environment data, c j represents the jth cluster center, c k represents the kth cluster center, ||x i -c j || represents the Euclidean distance from the i-th data point to the j-th cluster center, ||x i -c k || represents the Euclidean distance from the i-th data point to the k-th cluster center, m is the fuzzy coefficient, m is a real number greater than 1, which is used to control the fuzziness of the membership. The larger the m value, the more fuzzy the membership. ij Represents data point x i For cluster center c j The degree of membership.

[0046] S1023. Update each cluster center according to the membership degree of multiple historical meteorological environment data. The update formula of the cluster center is:

[0047]

[0048] Where N represents the total number of historical meteorological environment data.

[0049] S1024, judging whether the updated cluster centers have converged, and if so, outputting the final cluster centers and the membership matrix corresponding to the multiple historical meteorological environment data.

[0050] In this embodiment, judging whether the updated cluster center converges is judging whether the change of the cluster center in two consecutive iterations is less than a preset threshold.

[0051] In this embodiment, if the updated cluster centers have not converged, steps S22 to S23 are repeatedly performed until the updated cluster centers converge, and the final membership matrix of the cluster centers and data points is output.

[0052] In this embodiment, each row of the membership matrix represents a historical meteorological environment data, and each column represents the membership of a cluster (or category). For example, a 3x31 matrix represents 31 historical meteorological environment data, and each historical meteorological environment data has different memberships to three clusters.

[0053] S1025. According to the membership matrix of the historical meteorological and environmental data, the cluster to which each historical meteorological and environmental data belongs is determined according to the maximum membership principle in the fuzzy set, and a plurality of meteorological clusters are obtained.

[0054] In this embodiment, for each row in the membership matrix (ie, each historical meteorological environment data), the maximum membership value in the row is found, and the column number (cluster number) of the maximum membership value is the cluster to which the historical meteorological environment data belongs.

[0055] According to the above steps, all rows of the membership matrix are traversed, and each meteorological environment data sample is classified into the cluster corresponding to its maximum membership. In this way, all meteorological environment data are divided into multiple meteorological clusters.

[0056] For example, if the maximum membership value of the first row is the second column (0.5858), then the historical meteorological environment data is classified as the second category.

[0057] In this embodiment, in step S103, the temporal convolutional network TCN is a neural network architecture for processing sequence data. The temporal convolutional network TCN is composed of a plurality of stacked convolutional layers, each of which includes a causal convolution module, a dilated convolution module and a residual connection module, and the output of each layer is used as the input of the next layer.

[0058] The causal convolution module is used to ensure the causality of the model when predicting time series. The network structure of the causal convolution module is expressed as:

[0059]

[0060] In the formula, y t represents the output at time point t, x t-τ represents the input at time point t-τ, w τ represents the weight of the convolution kernel, b is the bias term, f is the activation function, and T is the size of the convolution kernel.

[0061] The dilated convolution module is used to capture longer-distance dependencies by increasing the spacing between elements in the convolution kernel (dilation factor). The network structure of the dilated convolution module is expressed as:

[0062]

[0063] Where D is the expansion factor.

[0064] The residual connection module is used to alleviate the gradient vanishing problem in deep network training and improve the training efficiency of deep networks by directly adding the output of the previous layer to the output of the current layer. The network structure of the residual connection module is expressed as:

[0065] y t =f(y res +x t ) (5)

[0066] In the formula, y res represents the output of the residual connection from the previous layer, x t Represents the input of the current layer.

[0067] In this embodiment, when training the temporal convolutional network TCN, an optimization algorithm such as gradient descent can be used to adjust the network parameters to minimize the loss function. At the same time, a regularization term can be introduced to prevent overfitting.

[0068] In this embodiment, the temporal convolutional network TCN uses a one-dimensional convolution layer to extract local features of sequence data, which makes the temporal convolutional network TCN faster when processing long sequences. Secondly, the temporal convolutional network TCN introduces the concept of dilated convolution, which is a special convolution operation that increases the receptive field by inserting spatial intervals in the convolution kernel, allowing the network to capture a wider range of contextual information without increasing the number of parameters. This mechanism enables the temporal convolutional network TCN to recognize and utilize patterns and trends that are far apart in the sequence. In addition, the temporal convolutional network TCN adopts the design principle of causal convolution, which ensures that the model does not use future information when making predictions, which is crucial to maintaining the causality and rationality of the model. The temporal convolutional network TCN also includes residual connections, a technique that allows the network to learn residual functions, which can alleviate the problem of gradient disappearance in deep network training.

[0069] In this embodiment, the input meteorological environment data is feature extracted by a temporal convolutional network (TCN). Each convolutional layer in the temporal convolutional network (TCN) outputs a set of feature maps, and the output feature maps contain feature information of the input data at different time scales.

[0070] In this embodiment, in step S104, the improved Autoformer model includes an encoder module, a decoder module and an output layer, the decoder module receives the decomposition components output by the encoder module, and decodes and reconstructs the decomposition components, the output layer receives the output results of the decoder module, and outputs the prediction results, the encoder module includes multiple MOEDecomp sequence decomposition layers and autocorrelation mechanism layers, and the decoder module includes a cumulative structure for processing trend periodic elements and a stacked autocorrelation structure for processing seasonal elements.

[0071] In this embodiment, the output layer is a fully connected layer.

[0072] In this embodiment, multiple training samples are input into the improved Autoformer model for training to obtain a short-term power system load forecasting model, which specifically includes: inputting multiple training samples as long sequences into the encoder module to obtain the trend sequence and seasonal feature sequence generated by the encoder module; inputting the generated trend sequence and seasonal feature sequence into the decoder module to obtain the output result of the decoder module; inputting the output result of the decoder module into the output layer to obtain the prediction result output by the output layer; calculating the loss function according to the prediction result output by the output layer and the actual load data, and performing gradient calculation and updating on the parameters of the improved Autoformer model according to the loss function through the back propagation algorithm to obtain the short-term power system load forecasting model.

[0073] In this embodiment, the encoder module converts the input time series data into a series of latent variables, which will be used in the subsequent decoding and prediction processes.

[0074] In this embodiment, the encoder module generates a trend sequence and a seasonal feature sequence by capturing the long-term trend and seasonal characteristics in the sequence. The trend sequence represents the long-term trend in the load data; the seasonal feature sequence contains periodic changes related to seasons, dates (such as days of the week, holidays), etc.

[0075] Specifically, the MOEDecomp sequence decomposition layer is used to decompose the input training samples into trend components and seasonal components. The calculation formulas of the trend component and the seasonal component are respectively expressed as:

[0076]

[0077] Where Xtrend represents the trend component, Xseaonal represents the seasonal component, x represents the input training sample, F(x) is a set of average pooling filters, L(x) is the weight used to mix the trend, and the Softmax function is used to ensure that the sum of the weights is 1.

[0078] The autocorrelation mechanism layer mainly includes a period-based dependency discovery mechanism and a time-delay information aggregation mechanism. This mechanism utilizes the periodicity of the sequence, discovers the dependencies of similar subsequences in different periods, and aggregates the information of these subsequences to achieve sequence-level connection.

[0079] Specifically, the cycle-based dependency discovery mechanism aims to discover and utilize the dependencies between these similar subsequences to improve the prediction ability of the model. It evaluates the similarity of the sequence under different time delays by calculating the autocorrelation coefficient of the time series, and identifies the cycle segments with similar subprocesses based on the calculation results of the autocorrelation coefficient. The subsequences of these cycle segments at the same phase position show high similarity and can therefore be regarded as subsequences with dependencies.

[0080] Furthermore, after identifying similar subsequences with dependencies, Autoformer aggregates the information of these subsequences through a time-delayed information aggregation mechanism. First, the similar subsequences are aligned according to the estimated period length. After the information is aligned, the information is aggregated using a query, key, and value format similar to the self-attention mechanism. The aggregated subsequence information is obtained by calculating the similarity between the query vector (query) and the key vector (key), and weighting the aggregation value vector (value) according to the similarity.

[0081] In this embodiment, the decoder module gradually accumulates trend information according to the generated trend sequence, and integrates the seasonal feature sequence into the trend prediction to generate the prediction result. The output of the decoder module is the load prediction value for each time step.

[0082] In this embodiment, the decoder module includes a cumulative structure to handle trend periodic elements and a stacked autocorrelation structure to handle seasonal elements.

[0083] Specifically, the main task of the accumulation structure for processing trend periodic elements is to gradually extract and accumulate long-term stable trend periodic elements from the predicted intermediate hidden variables. Through this accumulation structure, the decoder module can gradually refine the trend prediction and eliminate interference information in order to more accurately capture the long-term development pattern in the sequence. Assuming there are M decoder module layers, each decoding layer will further process and accumulate based on the output of the previous layer and the output of the encoder module.

[0084] The main task of the stacked autocorrelation structure for handling seasonal elements is to capture and handle seasonal elements in time series. The stacked autocorrelation structure finds period-based dependencies by calculating the autocorrelation coefficients of the series and aggregates similar subsequences from the underlying period. The stacked autocorrelation structure identifies subsequences with similar periodic characteristics by calculating the autocorrelation coefficients of the series and aligns and aggregates these subsequences using the time-delayed aggregation block. This stacked structure enables the decoder module to capture the changes and periodic characteristics of seasonal patterns more comprehensively.

[0085] In this embodiment, multiple training samples are divided into a training set and a test set. When the improved Autoformer model is trained, the training set is input for training. After the training is completed, the test set is input for testing the trained model. If the model meets the requirements, the model is used as a short-term power system load forecasting model.

[0086] In this embodiment, the improved Autoformer model uses the MOEDecomp sequence decomposition module to utilize a set of averaging filters of different sizes to capture multiple trend features in the input signal, and combines these trend features through data-related weights to form a final trend representation, which greatly effectively processes long time series with complex periodic patterns and trend components, helps to reduce the impact of the unstable characteristics of meteorological and environmental data on the prediction results, thereby improving the prediction accuracy of the short-term power system load forecasting model.

[0087] See also Figure 2 As shown, the embodiment of the present invention also provides a short-term power load forecasting method, including:

[0088] S201, obtaining predicted meteorological environment data within the predicted time, and extracting features of the predicted meteorological environment data;

[0089] S202: Input the extracted features into a short-term power system load forecasting model to obtain short-term power system load forecasting data.

[0090] In this embodiment, before extracting the features of the predicted meteorological environment data, the method further includes: preprocessing the acquired predicted meteorological environment data.

[0091] Specifically, the acquired forecast meteorological environment data is preprocessed, including data missing value processing, normalization processing and outlier processing.

[0092] In this embodiment, the features of the predicted meteorological environment data are extracted through a temporal convolutional network (TCN).

[0093] In this embodiment, based on the predicted meteorological and environmental data within the prediction time, prediction is performed through the short-term power system load prediction model to obtain the predicted meteorological and environmental data within the prediction time. The short-term power system load prediction model is trained based on the improved Autoformer model. The improved Autoformer model uses a set of average filters of different sizes to capture multiple trend features in the input signal, and combines these trend features through data-related weights to form a final trend representation, which greatly effectively handles long time series with complex periodic patterns and trend components, helps to reduce the impact of the unstable characteristics of meteorological and environmental data on the prediction results, thereby improving the prediction accuracy of the short-term power system load prediction model, and then improving the prediction accuracy of the short-term power system load.

[0094] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. All equivalent structural changes made by using the contents of the present invention specification and drawings under the inventive concept of the present invention, or directly / indirectly applied in other related technical fields are included in the patent protection scope of the present invention.

Claims

1. A method for constructing a short-term power load forecasting model, characterized in that: The short-term power load forecasting model construction method comprises: Collect historical data of the power system, which includes historical load data of multiple power systems and corresponding historical meteorological environment data; Clustering multiple historical meteorological environment data to obtain multiple meteorological clusters; Inputting the historical meteorological environment data of the plurality of meteorological clusters into a temporal convolutional network (TCN) to extract the meteorological environment features of the meteorological clusters, and generating a plurality of training samples according to the extracted meteorological environment features and the corresponding historical load data of the power system; The multiple training samples are input into the improved Autoformer model for training to obtain a short-term power system load forecasting model.

2. The method for constructing a short-term power load forecasting model according to claim 1, characterized in that: When clustering multiple historical meteorological environment data to obtain multiple meteorological clusters, a fuzzy C-means clustering algorithm is used to cluster the multiple historical meteorological environment data to obtain multiple meteorological clusters. The method of clustering multiple historical meteorological environment data using a fuzzy C-means clustering algorithm to obtain multiple meteorological clusters includes: obtaining the number of clusters C, and randomly initializing C cluster centers; for each historical meteorological environment data, calculating its membership to each cluster center; updating each cluster center according to the membership of multiple historical meteorological environment data; judging whether the updated cluster center converges, and if so, outputting the final cluster center and the membership matrix corresponding to the multiple historical meteorological environment data; according to the membership matrix of the multiple historical meteorological environment data, determining the cluster to which each historical meteorological environment data belongs according to the maximum membership principle in the fuzzy set to obtain multiple meteorological clusters.

3. The method for constructing a short-term power load forecasting model according to claim 1, characterized in that: The temporal convolutional network (TCN) is composed of a plurality of stacked convolutional layers, each of which includes a causal convolution module, a dilated convolution module and a residual connection module. The causal convolution module is used to ensure the causality of the model when performing time series prediction, the dilated convolution module is used to capture longer-distance dependencies, and the residual connection module alleviates the gradient vanishing problem in deep network training by directly adding the output of the previous layer to the output of the current layer.

4. The method for constructing a short-term power load forecasting model according to claim 1, characterized in that: The improved Autoformer model includes an encoder module, a decoder module and an output layer. The decoder module receives the decomposition components output by the encoder module, and decodes and reconstructs the decomposition components. The output layer receives the output results of the decoder module and outputs the prediction results. The encoder module includes multiple MOEDecomp sequence decomposition layers and autocorrelation mechanism layers. The decoder module includes a cumulative structure for processing trend periodic elements and a stacked autocorrelation structure for processing seasonal elements.

5. The method for constructing a short-term power load forecasting model according to claim 4, characterized in that: The step of inputting the plurality of training samples into the improved Autoformer model for training to obtain a short-term power system load forecasting model comprises: Inputting a plurality of training samples as long sequences into the encoder module, and obtaining a trend sequence and a seasonal feature sequence generated by the encoder module; Inputting the generated trend sequence and seasonal feature sequence into the decoder module to obtain the output result of the decoder module; Inputting the output result of the decoder module into the output layer to obtain the prediction result output by the output layer; The loss function is calculated based on the prediction result output by the output layer and the actual load data, and the parameters of the improved Autoformer model are gradient calculated and updated according to the loss function through the back propagation algorithm to obtain a short-term power system load forecasting model.

6. The method for constructing a short-term power load forecasting model according to claim 4, characterized in that: The MOEDecomp sequence decomposition layer is used to decompose the input training samples into trend components and seasonal components. The calculation formulas of the trend component and the seasonal component are respectively expressed as: X trend =Soft max(L(x))*(F(x)) X seasonal =x-X trend Where Xtrend represents the trend component, Xseaonal represents the seasonal component, x represents the input training sample, F(x) is a set of average pooling filters, L(x) is the weight used to mix the trend, and the Softmax function is used to ensure that the sum of the weights is 1.

7. The method for constructing a short-term power load forecasting model according to claim 4, characterized in that: The autocorrelation mechanism layer includes a period-based dependency discovery mechanism and a delay information aggregation mechanism. The period-based dependency discovery mechanism is used to identify similar subsequences with dependencies in a time series, and the delay information aggregation mechanism is used to align and aggregate information of similar subsequences.

8. A short-term power load forecasting method, characterized in that: The short-term power load forecasting method comprises: Obtain the forecasted meteorological environment data within the forecast period, and extract the features of the forecasted meteorological environment data; The extracted features are input into a short-term power system load forecasting model to obtain short-term power system load forecasting data, wherein the short-term power system load forecasting model adopts the short-term power system load forecasting model as described in any one of claims 1-7.

9. The short-term power load forecasting method according to claim 8, characterized in that: Before extracting the features of the predicted meteorological environment data, the method also includes: preprocessing the acquired predicted meteorological environment data.

Images