Power load prediction method, device and electronic equipment of power system

Abstract

The application discloses a power load prediction method and device of a power system and electronic equipment. The method comprises the following steps: obtaining historical power load data of the power system, wherein the historical power load data comprises a plurality of power load time series, and different power load time series are used to represent time series corresponding to different variables affecting the power load; inputting the plurality of power load time series into a power load prediction model; and predicting the plurality of power load time series of the power system in a future time period by using the power load prediction model to obtain predicted power load data, wherein the predicted power load data is used to represent the predicted power load data of the power system in the future time period. The application solves the technical problem of low power load prediction accuracy in the related art.

Description

Technical Field

[0001] This invention relates to the field of power systems, and more specifically, to a method, apparatus, and electronic device for predicting power load in a power system. Background Technology

[0002] Electricity load forecasting, a fundamental task in the power sector, plays a crucial role in the stable and efficient operation of the power system. With the integration of large-scale distributed power sources and new loads such as electric vehicle charging stations, the volatility, time-varying nature, and randomness of electricity load have significantly increased, making electricity load forecasting more challenging. Traditional forecasting methods, due to their linear assumptions, struggle to capture the complex nonlinear characteristics of electricity load data, resulting in low forecasting accuracy.

[0003] In recent years, deep learning models have been used to predict power load in related technologies. Due to the powerful nonlinear relationship learning ability of deep learning models, they have shown significant advantages in handling long-sequence data and multi-feature situations. However, existing models still have limitations in handling the uncertainty, randomness and complex time correlation of power load data, resulting in insufficient mining of the time series characteristics of power load, which in turn leads to low accuracy of power load prediction in related technologies.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This invention provides a method, apparatus, and electronic device for predicting power load in a power system, in order to at least solve the technical problem of low accuracy in power load prediction in related technologies.

[0006] According to one aspect of the present invention, a power load forecasting method for a power system is provided, comprising: acquiring historical power load data of the power system, the historical power load data including multiple power load time series, wherein different power load time series are used to represent time series corresponding to different variables affecting the power load; inputting the multiple power load time series into a power load forecasting model, and using the power load forecasting model to forecast the multiple power load time series of the power system in a future time period to obtain predicted power load data, wherein the predicted power load data is used to represent the predicted power load data of the power system in the future time period.

[0007] Furthermore, the power load forecasting model includes a time series decomposition module, a feature extraction module, and a feature enhancement module. Multiple power load time series are input into the power load forecasting model, which then predicts the power system's power load over a future period to obtain target predicted power load data. This includes: using the time series decomposition module to decompose multiple power load time series into trend time series components and seasonal time series components, where the seasonal time series components exhibit periodic fluctuations, and the trend time series components exhibit long-term stable trends; using the feature extraction module to extract features from the seasonal time series components to obtain comprehensive time window load features; and using the feature enhancement module to enhance the comprehensive time window load features and the trend time series components to obtain predicted power load data.

[0008] Furthermore, the feature extraction module includes: a time window embedding layer, a horizontal time window analyzer, and a vertical time window analyzer. The feature extraction module extracts features from the seasonal time series components to obtain comprehensive time window load features. This includes: using the time window embedding layer to establish a time window load feature matrix for the first and second load points in the seasonal time series components, where the second load point represents the load point within the adjacent time window of the first load point, and the time window load feature matrix represents the time window load features at multiple times; using the horizontal time window analyzer to extract the time window load features at multiple times to obtain horizontal time window load features at multiple times; and using the vertical time window analyzer to extract the horizontal time window load features at multiple times to obtain comprehensive time window load features at multiple times.

[0009] Furthermore, the feature enhancement module is used to enhance the comprehensive time window load characteristics and trend time series components to obtain predicted power load data, including: using the feature enhancement module to enhance the comprehensive time window load characteristics to obtain the seasonal time series component output results; using the feature enhancement module to enhance the trend time series components to obtain the trend time series component output results; and determining the predicted power load data based on the trend time series component output results and the seasonal time series component output results.

[0010] Furthermore, the feature enhancement module includes a first convolutional layer and a second convolutional layer. The feature enhancement module enhances the comprehensive time window load features to obtain the seasonal time series component output, and enhances the trend time series components to obtain the trend time series component output. This includes: using the first convolutional layer to extract features from the trend time series component and the comprehensive time window load features respectively, obtaining the first trend feature and the first seasonal feature; and using the second convolutional layer to extract features from the trend time series component and the comprehensive time window load features respectively, obtaining the second trend feature and the second seasonal feature. The convolution kernel of the second convolutional layer... The size of the convolutional kernel is larger than that of the first convolutional layer; the first trend convolutional parameters are determined based on the first activation function, the first trend feature, and the second trend feature; the first seasonal convolutional parameters are determined based on the first activation function, the first seasonal feature, and the second seasonal feature; the second trend convolutional parameters are determined based on the second activation function, the second trend feature, and the first trend feature; the second seasonal convolutional parameters are determined based on the second activation function, the second seasonal feature, and the first seasonal feature; the trend time series component output is obtained based on the first trend convolutional parameters and the second trend convolutional parameters; the seasonal time series component output is obtained based on the first seasonal convolutional parameters and the second seasonal convolutional parameters.

[0011] Furthermore, a time window load feature matrix is ​​established for the first load point and the second load point in the seasonal time series component using a time window embedding layer. This includes: constructing a preset dimension spatial feature of the seasonal time series component based on the first load point, the second load point, and preset time window parameters, wherein the preset dimension spatial feature is used to represent the spatial features in the preset dimension; and performing a linear transformation on the preset dimension spatial feature to obtain the time window load feature matrix.

[0012] Furthermore, the horizontal time window analyzer includes: a local multilayer perceptron and a first normalization layer. The horizontal time window analyzer is used to analyze the local correlation of time window load characteristics at multiple time points to obtain horizontal time window load characteristics at multiple time points. This includes: using the local multilayer perceptron to extract features from the time window load characteristics at multiple time points to obtain first time window load characteristics at multiple time points; using the first normalization layer to normalize the first time window load characteristics at multiple time points to obtain second time window load characteristics at multiple time points; and transposing the second time window load characteristics at multiple time points to obtain horizontal time window load characteristics at multiple time points.

[0013] Furthermore, the longitudinal time window analyzer includes a global multilayer perceptron and a second normalization layer. The longitudinal time window analyzer extracts the load features of the horizontal time window at multiple time points to obtain the comprehensive load features of the time window at multiple time points. This includes: extracting features from the horizontal time window load features at multiple time points using the global multilayer perceptron to obtain the load features of the third time window; normalizing the load features of the third time window at multiple time points using the second normalization layer to obtain the load features of the fourth time window at multiple time points; and transposing the load features of the fourth time window at multiple time points to obtain the comprehensive load features of the time window at multiple time points.

[0014] According to another aspect of the present invention, a power load forecasting device for a power system is also provided, comprising: an acquisition module for acquiring historical power load data of the power system, the historical power load data including multiple power load time series, wherein different power load time series are used to represent time series corresponding to different variables affecting the power load; and a forecasting module for inputting the multiple power load time series into a power load forecasting model, and using the power load forecasting model to forecast the multiple power load time series of the power system in a future time period to obtain predicted power load data, wherein the predicted power load data is used to represent the predicted power load data of the power system in the future time period.

[0015] According to another aspect of the present invention, an electronic device is also provided, comprising: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods of various embodiments of the present invention during runtime.

[0016] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is executed, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of the present invention.

[0017] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of the present invention.

[0018] According to another aspect of the present invention, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods of various embodiments of the present invention.

[0019] In this embodiment of the invention, firstly, historical power load data of the power system is acquired. Then, this historical power load data is input into a power load prediction model. The power load prediction model predicts multiple power load time series of the power system in the future time period, obtaining predicted power load data. It is noteworthy that the historical power load data contains multiple power load time series, each corresponding to a variable affecting power load prediction. By considering multiple variables affecting power load and designing the power load model to capture both global and local correlations between power load time series, the goal of in-depth mining of power load characteristics is achieved, thereby improving prediction accuracy and solving the technical problem of low power load prediction accuracy in related technologies. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0021] Figure 1 This is a flowchart of a power load forecasting method for a power system according to an embodiment of the present invention;

[0022] Figure 2 This is a flowchart of an optional power load forecasting method for a power system according to an embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram of a power load prediction device for a power system according to an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] According to an embodiment of the present invention, an embodiment of a power load forecasting method for a power system is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0027] Figure 1 This is a flowchart of a power load forecasting method for a power system according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0028] Step S102: Obtain historical power load data of the power system. The historical power load data includes multiple power load time series, wherein different power load time series are used to represent the time series corresponding to different variables affecting the power load.

[0029] The historical power load data mentioned above describes the power consumption records of the power system over a certain period of time in the past. The historical load data is presented in the form of a time series, and the load data at each point in time reflects the demand of the power system at that moment. The historical power load data includes multiple power load time series. The functions of historical power load data can include, but are not limited to, power load forecasting, analyzing power demand patterns, and formulating electricity pricing strategies.

[0030] The power load time series in the above steps is used to describe a series of load values ​​arranged in chronological order in historical power load data. Different power load time series represent the changes in power load over time corresponding to different variables. Among them, different variables are used to describe multiple variables affecting power load. These multiple variables may include, but are not limited to, temperature, humidity, power consumption, commercial electricity consumption, and electricity price. Power load time series may include, but are not limited to, temperature-related load series, humidity-related load series, and power consumption-related time series. The specific power load time series needs to be determined based on the specific influencing variables, and is not limited here.

[0031] The methods for obtaining the information in the above steps may include, but are not limited to, the following:

[0032] The first method involves using a Supervisory Control and Data Acquisition (SCADA) system to monitor and record the operating status of the power network in real time, thereby obtaining historical power load data.

[0033] The second method involves obtaining historical electricity load data by acquiring user-side electricity consumption information through smart meters and obtaining historical electricity price information from publicly available data from power companies.

[0034] The above methods for obtaining historical power load data are only examples. The specific methods for obtaining the data need to be determined based on the actual situation of the power system, and are not limited here.

[0035] In one optional embodiment, historical power load data of the power system is obtained through various channels such as smart meters and publicly available electricity prices from power companies. This historical power load data contains multiple power load time series, each representing the changes over time in different variables affecting power load. For example, temperature-related load series reflect the impact of ambient temperature fluctuations on electricity demand, while social activity-related load series record patterns of electricity consumption changes caused by special events or holidays. Through training and analysis of multiple power load time series, the periodicity, seasonality, and event-driven characteristics of historical power load data can be understood, facilitating subsequent power load forecasting based on this historical data.

[0036] In one optional embodiment, historical power load data undergoes a reversible instance normalization module, which removes non-stationary information from the data and addresses the data distribution drift problem. Reversible instance normalization includes normalization and inverse normalization. First, the initial input data is normalized for each input power load time series x. (i) ∈R 1×M (i = 1, 2, ..., N), at time t, are instance-standardized using their own mean and variance, and the calculation formula is as follows:

[0037]

[0038] Using the mean and variance to analyze the input power load time series x (i) The standardization calculation formula is as follows:

[0039]

[0040] In the formula, This represents the time series representation of the input power load at time t, where M represents x. (i) The length of the power load time series, μ (i) and σ (i) They represent the values ​​for x respectively (i) The mean and variance are calculated, where α and β are the learnable parameters of the model, and ε is the random error. For time series x (i) The standardized result. After standardization, the power load time series is represented as follows: Its corresponding output is

[0041] Accordingly, the output of the power load forecasting model, i.e., the forecasted power load data, was inversely standardized, and the calculation process is as follows:

[0042]

[0043] In the formula, The output of the model before inverse standardization. Input variables The corresponding output results.

[0044] Step S104: Input multiple power load time series into the power load prediction model, and use the power load prediction model to predict multiple power load time series of the power system in the future time period to obtain predicted power load data, wherein the predicted power load data is used to represent the predicted power load data of the power system in the future time period.

[0045] The power load forecasting model mentioned above is a deep learning model used to analyze patterns in historical power load data and predict future power demand. The type of power load forecasting model may include, but is not limited to, deep learning models including Long Short-Term Memory Networks and Convolutional Neural Networks. The specific power load forecasting model needs to be determined according to the nature of the power load data and actual needs, and is not limited here. The functions of the power load forecasting model may include, but are not limited to, power load forecasting, demand response, and resource planning.

[0046] The future time period mentioned in the above steps describes the time range in which the power load forecasting model will make predictions. The type of future time period can include, but is not limited to, the next few hours, days, weeks or months. The specific future time period depends on the needs of the forecasting task and the capabilities of the model, and is not limited here.

[0047] The predicted power load data in the above steps is a time series describing the future power system demand predicted by the power load forecasting model. The role of the predicted power load data can include, but is not limited to, operation optimization, economic decision-making, and power system planning based on the predicted power load data.

[0048] In one optional embodiment, multiple power load time series are input into a power load prediction model. By training on multiple historical power load time series, the power load prediction model can analyze the periodic characteristics of historical power load data. Based on the periodic characteristics of the acquired historical power load data, the power load prediction model can predict the power load time series for future periods, obtain the predicted power load data for future periods, and then improve the power resource allocation and economic dispatch strategies of the power market based on the predicted power load data, providing strong support for the efficient, stable, and economical operation of the power system.

[0049] In this embodiment of the invention, firstly, historical power load data of the power system is acquired. Then, this historical power load data is input into a power load prediction model. The power load prediction model predicts multiple power load time series of the power system in the future time period, obtaining predicted power load data. It is noteworthy that the historical power load data contains multiple power load time series, each corresponding to a variable affecting power load prediction. By considering multiple variables affecting power load and designing the power load model to capture both global and local correlations between power load time series, the goal of in-depth mining of power load characteristics is achieved, thereby improving prediction accuracy and solving the technical problem of low power load prediction accuracy in related technologies.

[0050] Optionally, the power load forecasting model includes: a time series decomposition module, a feature extraction module, and a feature enhancement module. Multiple power load time series are input into the power load forecasting model, which then forecasts the power system's multiple power load time series over future periods to obtain target predicted power load data. This includes: using the time series decomposition module to decompose multiple power load time series into trend time series components and seasonal time series components, where the seasonal time series components exhibit periodic fluctuations, and the trend time series components exhibit long-term stable trends; using the feature extraction module to extract features from the seasonal time series components to obtain comprehensive time window load features; and using the feature enhancement module to enhance the comprehensive time window load features and the trend time series components to obtain predicted power load data.

[0051] The aforementioned time series decomposition module describes a technique for decomposing the original power load time series into trend time series components and seasonal time series components. The types of time series decomposition can include, but are not limited to, average pooling operations in deep learning models, wavelet transform decomposition in signal processing, and Fourier transform decomposition. The role of the time series decomposition module can include, but is not limited to, enabling the model to better understand and process the different characteristics in the power load data through decomposition, providing clear signal components for subsequent feature extraction and prediction, thereby improving the accuracy and stability of prediction.

[0052] The aforementioned feature extraction module describes the technique for identifying and extracting key features from the decomposed seasonal time series components. The type of feature extraction module may include, but is not limited to, convolutional neural networks, long short-term memory networks, etc. The specific feature extraction module needs to be determined according to the model structure design, and is not limited here. The role of the feature extraction module may include, but is not limited to, extracting key information from the original data, providing more focused and meaningful input for the prediction model, and helping to enhance the model's expressive power and prediction accuracy.

[0053] The aforementioned feature enhancement module describes the process of further strengthening feature representation based on feature extraction. The type of feature enhancement module may include, but is not limited to, improved convolutional neural networks, residual connections, etc. The specific feature enhancement module needs to be determined according to the model design and feature type, and is not limited here. The role of the feature enhancement module may include, but is not limited to, helping the model to more comprehensively understand the inherent laws of power load data and improving prediction accuracy.

[0054] The aforementioned comprehensive time window load characteristics are used to describe the features extracted from seasonal time series components and take into account the correlation between load data within the time window.

[0055] In one optional embodiment, multiple power load time series are input into a power load prediction model that includes a time series decomposition module, a feature extraction module, and a feature enhancement module. The power load prediction model first uses the time series decomposition module to decompose the multiple power load time series into trend time series components and seasonal time series components, enabling the model to effectively identify trends and seasonal variations in the power load data. Then, features are extracted based on the characteristics of different time series components, ultimately obtaining highly accurate target power load prediction data. This time series processing method of the power load prediction model can effectively improve the model's power load prediction accuracy and has significant practical application value for the optimized operation of power systems, economic dispatch of the power market, and intelligent electricity management for users.

[0056] In one optional embodiment, the time series decomposition module decomposes multiple power load time series into trend time series components and seasonal time series components, and then extracts features from the trend time series components and seasonal time series components respectively. First, the moving average method is used to calculate the average of the power load time series over a period of time to smooth the periodic terms and highlight the long-term trend of the power load time series. After obtaining the trend time series component of the power load time series, the seasonal time series component is obtained by subtracting the trend time series component from the input power load time series. The formula for calculating the moving average is as follows:

[0057]

[0058] In the formula, A t n represents the moving average at time point t. LS B represents the sliding window size. i This represents the original power load time series at the i-th time step.

[0059] For a multivariate time series at time t Its decomposition formula is as follows:

[0060]

[0061] X trend and X season These represent the trend time series component and the seasonal time series component, respectively, decomposed by the time series decomposition module. The trend time series component contains long-term and coarse information about the time series, while the seasonal time series component reflects short-term and detailed information about the time series. After time series decomposition, it is easier for the model to process dynamically changing data.

[0062] Optionally, the feature extraction module includes: a time window embedding layer, a horizontal time window analyzer, and a vertical time window analyzer. The feature extraction module extracts features from the seasonal time series components to obtain comprehensive time window load features. This includes: using the time window embedding layer to establish a time window load feature matrix for the first and second load points in the seasonal time series components, where the second load point represents the load point within the adjacent time window of the first load point, and the time window load feature matrix represents the time window load features at multiple times; using the horizontal time window analyzer to extract the time window load features at multiple times to obtain horizontal time window load features at multiple times; and using the vertical time window analyzer to extract the horizontal time window load features at multiple times to obtain comprehensive time window load features at multiple times.

[0063] The aforementioned time window embedding layer is used to describe the network layer in the feature extraction module used to construct the time window load feature matrix. The size of the time window can be, but is not limited to, 6, 12, 18, 24, etc. The specific size of the time window needs to be determined according to the sampling frequency of the data, which is not limited here. The function of the time window embedding layer can be, but is not limited to, associating local information of the load data with the context, constructing the time window load feature matrix, and constructing a high-dimensional feature space.

[0064] The aforementioned lateral time window analyzer is a network layer used to describe the local correlation of the load feature matrix for extracting time windows. The role of the lateral time window analyzer may include, but is not limited to, helping the model to gain a deeper understanding of short-term fluctuations and local patterns of power load.

[0065] The aforementioned longitudinal time window analyzer is used to describe the global correlation between features of the lateral time window. The role of the longitudinal time window analyzer may include, but is not limited to, helping the model identify long-term trends and long-distance dependencies of power load.

[0066] The first load point mentioned above is used to describe the load value at a specific point in time in a time series.

[0067] The aforementioned horizontal time window load characteristics are used to describe the features extracted from the time window load characteristic matrix by the horizontal time window analyzer. The horizontal time window load characteristics reflect the interrelationship of load points within the same time window.

[0068] In one optional embodiment, a feature extraction module for processing seasonal time series components in the power load forecasting model is constructed using a time window embedding layer, a horizontal time window analyzer, and a vertical time window analyzer. Then, the time window embedding layer is used to construct a time window load feature matrix, the horizontal time window analyzer is used to mine the local correlations of the time window load feature matrix to obtain horizontal time window features, and the vertical time window analyzer is used to mine the global correlations between the horizontal time window features to obtain comprehensive time window features. This feature extraction structure enables the model to capture both global and local correlations in power load data, thereby improving the accuracy of power load forecasting.

[0069] Optionally, the feature enhancement module is used to enhance the comprehensive time window load characteristics and trend time series components to obtain predicted power load data, including: using the feature enhancement module to enhance the comprehensive time window load characteristics to obtain the seasonal time series component output result; using the feature enhancement module to enhance the trend time series components to obtain the trend time series component output result; and determining the predicted power load data based on the trend time series component output result and the seasonal time series component output result.

[0070] The above-mentioned seasonal time series component output results are used to describe the feature results extracted and enhanced by the feature enhancement module from the seasonal time series components. The seasonal time series component output results reflect the periodicity and short-term fluctuation information in the load data, which can help the model identify and predict seasonal or periodic electricity demand.

[0071] The aforementioned trend time series component output results are used to describe the feature enhancement results obtained by the feature enhancement module when the model processes long-term trend information. The trend time series component output results reflect the overall trend of electricity demand over time, which can help the model understand and predict long-term electricity load growth or decline.

[0072] In one optional embodiment, the power load forecasting model, by processing seasonal time components and trend time series components separately, can capture short-term fluctuations and long-term trends in power load from different perspectives. During the model's forecasting phase, the outputs of the seasonal time components and the trend time series components are aggregated using methods such as summation and weighted averaging to generate the final predicted power load data. This aggregation operation integrates the model's understanding of short-term changes and long-term trends, providing comprehensive power demand forecasts. This helps power companies conduct accurate power resource planning and dispatching, while also providing users with more refined electricity consumption guidance, promoting the economical and efficient operation of the power system.

[0073] Optionally, the feature enhancement module includes: a first convolutional layer and a second convolutional layer. The feature enhancement module enhances the comprehensive time window load features to obtain the seasonal time series component output, and enhances the trend time series components to obtain the trend time series component output. This includes: using the first convolutional layer to extract features from the trend time series component and the comprehensive time window load features respectively, obtaining a first trend feature and a first seasonal feature; and using the second convolutional layer to extract features from the trend time series component and the comprehensive time window load features respectively, obtaining a second trend feature and a second seasonal feature. The kernel size of the second convolutional layer is... The kernel size is larger than that of the first convolutional layer; the first trend convolution parameters are determined based on the first activation function, the first trend feature, and the second trend feature; the first seasonal convolution parameters are determined based on the first activation function, the first seasonal feature, and the second seasonal feature; the second trend convolution parameters are determined based on the second activation function, the second trend feature, and the first trend feature; the second seasonal convolution parameters are determined based on the second activation function, the second seasonal feature, and the first seasonal feature; the trend time series component output is obtained based on the first trend convolution parameters and the second trend convolution parameters; the seasonal time series component output is obtained based on the first seasonal convolution parameters and the second seasonal convolution parameters.

[0074] The first and second convolutional layers mentioned above are convolutional network layers used for feature enhancement in the feature enhancement module. The kernel size of the second convolutional layer is larger than that of the first convolutional layer. The kernel size can be, but is not limited to, 1×1, 3×3, 5×5, etc. The specific kernel size needs to be determined according to the network structure and feature structure. It is not limited here. In one optional case, the kernel size of the first convolutional layer can be set to 1×1 and the kernel size of the second convolutional layer can be set to 3×3.

[0075] The first trend feature and the second trend feature mentioned above are used to describe the results of the model learning trend information. When the feature enhancement module processes the trend time series components, the first trend feature and the second trend feature extracted by the first convolutional layer and the second convolutional layer, respectively, are the results of the model learning trend information.

[0076] The first and second seasonal features mentioned above are used to describe the results of learning seasonal information when the feature enhancement module processes the comprehensive time window load features and extracts the first and second seasonal features through the first and second convolutional layers, respectively.

[0077] The first activation function mentioned above describes the activation function located after the first convolutional layer. It is used to help the model learn more complex feature representations. The first activation function may include, but is not limited to, the Gaussian Error Linear Unit (GELU), the Rectified Linear Unit (ReLU), the Sigmoid activation function, etc. The specific type of the first activation function can be selected according to the function of the activation function, and is not limited here.

[0078] The second activation function mentioned above describes the activation function located after the second convolutional layer. It is used to help the model learn more complex feature representations. The second activation function may include, but is not limited to, the Gaussian Error Linear Unit (GELU), the Rectified Linear Unit (ReLU), the Sigmoid, etc. The specific type of second activation function can be selected according to the function of the activation function, and is not limited here.

[0079] The first trend convolution parameter mentioned above is used to describe the feature obtained by multiplying the first trend feature by the second trend feature after activating the first trend feature with the first activation function.

[0080] The first seasonal convolution parameter mentioned above is used to describe the feature obtained by multiplying the first seasonal feature with the second seasonal feature after activating the first seasonal feature using the first activation function.

[0081] The aforementioned second-trend convolution parameter is used to describe the feature obtained by multiplying the second-trend feature by the first-trend feature after activating the second-trend feature with the second activation function.

[0082] The second seasonal convolution parameters mentioned above are used to describe the features obtained by activating the second seasonal features with the second activation function and then multiplying them with the first seasonal features.

[0083] In one optional embodiment, first trend features, first seasonal features, second trend features, and second seasonal features are extracted through a first convolutional layer and a second convolutional layer. Then, an activation function is used to further determine the convolution parameters, enabling the feature enhancement module to effectively learn key information from the electricity load data. This multi-scale feature extraction method, combining nonlinear activation functions and adaptive convolution parameter learning, significantly improves the model's performance in handling electricity load forecasting tasks. It allows the model to more comprehensively understand the temporal and periodic characteristics of electricity load, thereby producing more accurate load forecasting results.

[0084] In one optional embodiment, two convolutional layers of different sizes are used to extract features from the time series, and the convolutional layers are then reused through aggregation. The first convolutional layer uses fine-grained kernels to capture the modalities specific to the time series, and its calculation formula is as follows:

[0085]

[0086] In the formula, σ(·) represents the first activation function, and K (i) f represents the input features. conv1 (·) and f conv2 (·) represent the first and second convolutional layers, respectively. This represents the dot product operation of matrices. This indicates that the output of the first convolutional layer can be either the first trend convolutional parameter or the first seasonal convolutional parameter.

[0087] The second convolutional layer uses a coarse-grained convolutional kernel, primarily to capture long-term correlations hidden in the time series. The calculation formula is as follows:

[0088]

[0089] In the formula, σ(·) represents the second activation function, and K (i) f represents the input features. conv1 (·) and f conv2 (·) represent the first and second convolutional layers, respectively. This represents the dot product operation of matrices. This indicates that the output of the second convolutional layer can be either the second trend convolutional parameter or the second seasonal convolutional parameter.

[0090] The feature outputs from the two convolutional layers are summed and then fed into the last convolutional layer for processing. The output is as follows:

[0091]

[0092] In the formula, f conv3 (.) indicates the third convolutional layer. These represent the outputs of the first, second, and third convolutional layers, respectively.

[0093] After feature extraction by the feature enhancement module, the trend time series component and the seasonal time series component are predicted using a fully connected layer. Finally, the predicted power load data is obtained by summation, as shown in the following formula:

[0094]

[0095] In the formula, and These represent the output results of the trend time series component and the seasonal time series component, respectively. This is the predicted power load data after superimposing the two components. W trend and b trend W represents the weights and bias parameters of the fully connected layer for the trend time series components. season and b season These are the weights and bias parameters for the seasonal time series components.

[0096] Optionally, a time window load feature matrix is ​​established for the first load point and the second load point in the seasonal time series component using a time window embedding layer, including: constructing a preset dimension spatial feature of the seasonal time series component based on the first load point, the second load point and preset time window parameters, wherein the preset dimension spatial feature is used to represent the spatial feature in the preset dimension; and performing a linear transformation on the preset dimension spatial feature to obtain the time window load feature matrix.

[0097] The aforementioned preset time window parameter describes the size λ of the time window required to construct the preset dimension spatial features based on the first load point. The size of the time window can be, but is not limited to, 6, 12, 18, 24, etc. The specific size of the time window needs to be determined according to the data sampling frequency, which is not limited here. The function of the preset time window parameter can include, but is not limited to, adjusting the ability of the model to capture local and periodic features by adjusting the time window parameter.

[0098] In one optional embodiment, by determining a first load point and preset time window parameters, the model can construct spatial features representing the load within a specific time scale when processing seasonal time series components. Then, by performing a linear transformation on these preset-dimensional spatial features, the information is mapped to a higher-dimensional representation space that is more conducive to feature extraction and analysis, thereby obtaining a time window load feature matrix. The time window load feature matrix contains comprehensive information about the load point and its neighboring points, providing structured input for subsequent horizontal and vertical time window analyses.

[0099] In one optional embodiment, it is assumed that the multiple power load time series are represented in the decomposed form as follows: A time window feature is constructed using the load value at each time point m (i.e., the first load point) as the center, with a left and right length of λ. For the load values ​​at the first and last two times of the time series, a copy operation is performed when constructing the time window feature, resulting in a modified time series.

[0100] The load value x at time m m Construct an eigenvector v of length 2λ+1 centered at the center. m The formula is as follows:

[0101] v m =[x m-λ ,…,x m-1 ,x m ,x m+1 ,…,x m+λ ]∈R 1×(2λ+1) ;

[0102] In the formula, v m Let λ represent the feature vector, λ represent the size of the time window, and m represent the load value at time m.

[0103] Based on the above formula, the i-th seasonal component can be obtained. The high-dimensional feature space of is given by the following formula:

[0104] V (i) =[v1,v2,…,v M ] T ∈R M×(2λ+1) ;

[0105] In the formula, V (i) For the i-th seasonal component The high-dimensional feature space, where T represents the transpose operation and M represents the M seasonal components.

[0106] Next, a linear function is used to transform the feature space of the time window. This process allows us to obtain the global transient features of the load in a higher-dimensional space. The linear transformation formula is as follows:

[0107] u m =v m ·W+b;

[0108] In the formula, u m For v m The feature representation after linear transformation, W∈R (2λ+1)×D Let b be the weight matrix, and b ∈ R. 1×D Here, is the bias parameter, D is the dimension of the time window features after linear transformation, and the symbol · represents the dot product of matrices. All features after linear transformation are:

[0109] U (i) =[u1,u2,…,u M ] T ∈R M×D ;

[0110] U (i) For the i-th seasonal component The representation of the characteristics after linear transformation is the integrated time window load characteristics, where T represents the transpose operation.

[0111] Optionally, the horizontal time window analyzer includes: a local multilayer perceptron and a first normalization layer. The horizontal time window analyzer is used to analyze the local correlation of time window load characteristics at multiple time points to obtain horizontal time window load characteristics at multiple time points. This includes: using the local multilayer perceptron to extract features from the time window load characteristics at multiple time points to obtain first time window load characteristics at multiple time points; using the first normalization layer to normalize the first time window load characteristics at multiple time points to obtain second time window load characteristics at multiple time points; and transposing the second time window load characteristics at multiple time points to obtain horizontal time window load characteristics at multiple time points.

[0112] The aforementioned Local Multilayer Perceptron is a network layer used to describe local features for extracting load features within a time window. The role of the Local Multilayer Perceptron may include, but is not limited to, extracting the first time window load features from the time window load features, that is, capturing the local correlation of load data within a specific time window, which helps the model understand short-term trends and cyclical patterns, and improves prediction accuracy.

[0113] The first time window load feature mentioned above is used to describe the output of the local multilayer perceptron after extracting the load features of the time window. The first time window load feature represents the local feature representation obtained by the neural network processing of the load data at each time point within the time window.

[0114] The first normalization layer mentioned above is used to describe the network layer that normalizes the load characteristics of the first time window so that the data meets the normalization requirements. The types of the first normalization layer may include, but are not limited to, batch normalization (BN), layer normalization (LN), and instance normalization (IN). The specific first normalization layer needs to be determined according to the scenario requirements and model structure, and is not limited here.

[0115] The aforementioned second time window load feature is used to describe the result after the first time window load feature has been standardized by the first standardization layer. The second time window load feature has a uniform scale and distribution. The role of the second time window load feature includes, but is not limited to, preserving the feature information extracted by the local multilayer perceptron, ensuring the statistical consistency of the data, and providing standardized input for subsequent analysis and prediction tasks.

[0116] In one optional embodiment, the lateral time window analyzer extracts the first time window load features within each time window through a local multilayer perceptron, then processes them through a first normalization layer to obtain the second time window load features, and finally generates the lateral time window load features through a transpose operation. This series of steps together constitutes a complete process for capturing the local correlation of power load, providing strong support for power load forecasting.

[0117] In one optional embodiment, the process by which the lateral time window analyzer performs local correlation analysis on the time window load characteristics at multiple time points is as follows:

[0118] First, at time m∈[1,M], the load features of the time window are extracted using a local multilayer perceptron, and the calculation formula is as follows:

[0119]

[0120] In the formula, Ω (i) ∈R M×D The output of the local multilayer perceptron, i.e., the load characteristics of the first time window, is f. LocMLP (·) represents the expression for a local multilayer perceptron. The characteristic matrix represents the load characteristics within a time window at time m.

[0121] The output Ω of the local multilayer perceptron (i) Standardization will be performed using the first standardization layer, and its calculation formula is as follows:

[0122]

[0123] In the formula, The output of the transverse time window analyzer, i.e., the load characteristic of the second time window, is represented by f. norm (·) represents the function expression for the first normalized layer. Second time window load characteristics. The horizontal time window load characteristics are obtained by matrix transpose, and the calculation formula is as follows:

[0124]

[0125] In the formula, L (i) This indicates the load characteristics within a horizontal time window. This indicates the load characteristics of the second time window, and T represents the transpose operation.

[0126] Optionally, the longitudinal time window analyzer includes: a global multilayer perceptron and a second normalization layer. The longitudinal time window analyzer is used to extract the load features of the horizontal time window at multiple time points to obtain the comprehensive load features of the time window at multiple time points. This includes: using the global multilayer perceptron to extract the load features of the horizontal time window at multiple time points to obtain the load features of the third time window; using the second normalization layer to normalize the load features of the third time window at multiple time points to obtain the load features of the fourth time window at multiple time points; and transposing the load features of the fourth time window at multiple time points to obtain the comprehensive load features of the time window at multiple time points.

[0127] The aforementioned Global Multilayer Perceptron (GLMP) is a network layer used to describe global patterns and correlations for processing and extracting features of long-sequence data. The role of the GLMP can include, but is not limited to, capturing and understanding the global correlations hidden in load data between different time windows, thereby understanding the macro trends and periodic patterns of power load.

[0128] The aforementioned third time window load feature is used to describe the output of the global multilayer perceptron after extracting features from the horizontal time window load feature. The third time window load feature can be used to represent the comprehensive feature representation of load data in different time windows, including long-range dependencies and global pattern information.

[0129] The second normalization layer mentioned above is used to describe the normalization process of the load characteristics of the third time window so that the data meets the normalization requirements. The types of the second normalization layer may include, but are not limited to, batch normalization (BN), layer normalization (LN), and instance normalization (IN). The specific second normalization layer needs to be determined according to the scenario requirements and model structure, and is not limited here.

[0130] The aforementioned fourth time window load feature is used to describe the output of the second standardization layer after processing the third time window load feature. The role of the fourth time window load feature may include, but is not limited to, providing standardized and structured input for subsequent prediction and analysis steps, ensuring that the model can extract effective information from long-range dependencies.

[0131] In one optional embodiment, the longitudinal time window analyzer consists of a global multilayer perceptron and a second normalization layer. During feature extraction, the global multilayer perceptron first performs deep feature extraction on the load features of the transverse time windows generated by the transverse time window analyzer at multiple time points, generating third time window load features. This process reveals the global correlation between features of different time windows, thereby capturing the long-term trend and periodic changes in power load. Subsequently, the second normalization layer normalizes the third time window load features, transforming them into fourth time window load features. This ensures the consistency and comparability of these features across different time windows, improving the training stability and prediction accuracy of the model.

[0132] In one optional embodiment, the longitudinal time window analyzer performs a global correlation analysis on the load characteristics of the lateral time window at multiple time points as follows:

[0133] For each dimension d∈[1,2,...,D] of the horizontal time window load feature, the output calculation formula after processing by the global multilayer perceptron is as follows:

[0134]

[0135] In the formula, f represents the load characteristics of the horizontal time window. GLoMLP (.) represents the expression for the global multilayer perceptron. This represents the output of the global multilayer perceptron, i.e., the load characteristics of the third time window.

[0136] After performing the above transformation on the features of each dimension, the output Γ of the global multilayer perceptron is obtained. (i) ∈R D×M Next, standardization is performed using the second standardization layer, and the calculation formula is as follows:

[0137]

[0138] In the formula, f represents the load characteristics of the fourth time window. norm (·) represents the function expression of the second normalization layer, Γ d This indicates the load characteristics of the third time window.

[0139] Output of longitudinal time window analyzer The matrix transpose will be used as the output of the longitudinal time window analyzer, and its output calculation formula is as follows:

[0140]

[0141] In the formula, G represents the load characteristics of the fourth time window. (i)This indicates the overall load characteristics within a time window.

[0142] In one alternative embodiment, Figure 2 This is a flowchart of an optional power system load forecasting method according to an embodiment of the present invention, such as... Figure 2 As shown, this method first decomposes the input historical power load data into a power load time series, obtaining seasonal time series components and trend time series components. Then, the seasonal time series components are input into a feature extraction module containing a time window embedding layer, a horizontal time window analyzer, and a vertical time window analyzer for feature extraction. The extracted comprehensive time window features are then subjected to dimensionality reduction and input into a feature enhancement module. Simultaneously, the trend time series components are input into the feature enhancement module. The results of feature enhancement for the two time series components are summed after passing through a fully connected layer to obtain the output result, which is the predicted power load data.

[0143] According to another aspect of the present invention, a power load forecasting device for a power system is also provided. This device can execute the power load forecasting method for a power system provided in the above embodiments. The specific implementation and preferred application scenarios are the same as those in the above embodiments, and will not be described in detail here.

[0144] Figure 3 This is a schematic diagram of a power load forecasting device for a power system according to an embodiment of the present invention, as shown below. Figure 3 As shown, the device includes: an acquisition module 302, used to acquire historical power load data of the power system, the historical power load data including multiple power load time series, wherein different power load time series are used to represent the time series corresponding to different variables affecting the power load; and a prediction module 304, used to input multiple power load time series into a power load prediction model, and use the power load prediction model to predict multiple power load time series of the power system in the future time period to obtain predicted power load data, wherein the predicted power load data is used to represent the predicted power load data of the power system in the future time period.

[0145] Optionally, the power load forecasting model includes: a time series decomposition module, a feature extraction module, and a feature enhancement module. The forecasting module includes: a decomposition unit, used to decompose multiple power load time series using the time series decomposition module to obtain trend time series components and seasonal time series components, wherein the seasonal time series components are components exhibiting periodic fluctuations among multiple power load time series, and the trend time series components are components exhibiting long-term stable changing trends among multiple power load time series; a feature extraction unit, used to extract features from the seasonal time series components using the feature extraction module to obtain comprehensive time window load features; and a feature enhancement unit, used to enhance the comprehensive time window load features and trend time series components using the feature enhancement module to obtain predicted power load data.

[0146] Optionally, the feature extraction module includes: a time window embedding layer, a horizontal time window analyzer, and a vertical time window analyzer. The feature extraction unit includes: an establishment subunit, used to establish a time window load feature matrix for the first load point and the second load point in the seasonal time series component using the time window embedding layer, wherein the second load point is used to represent the load point in the adjacent time window of the first load point, and the time window load feature matrix is ​​used to represent the time window load features at multiple times; a first feature extraction subunit, used to extract the time window load features at multiple times using the horizontal time window analyzer to obtain the horizontal time window load features at multiple times; and a second feature extraction subunit, used to extract the horizontal time window load features at multiple times using the vertical time window analyzer to obtain the comprehensive time window load features at multiple times.

[0147] Optionally, the feature enhancement unit includes: a first feature enhancement subunit, used to enhance the load characteristics of the comprehensive time window using the feature enhancement module to obtain the seasonal time series component output result; a second feature enhancement subunit, used to enhance the trend time series component using the feature enhancement module to obtain the trend time series component output result; and a determination subunit, used to determine the predicted power load data based on the trend time series component output result and the seasonal time series component output result.

[0148] Optionally, the feature enhancement module includes: a first convolutional layer, a second convolutional layer, a first feature enhancement subunit, and a second feature enhancement subunit, comprising: using the first convolutional layer to extract features from the trend time series component and the comprehensive time window load feature respectively, to obtain a first trend feature and a first seasonal feature; using the second convolutional layer to extract features from the trend time series component and the comprehensive time window load feature respectively, to obtain a second trend feature and a second seasonal feature, wherein the convolution kernel size of the second convolutional layer is larger than the convolution kernel size of the first convolutional layer; determining first trend convolution parameters based on a first activation function, the first trend feature, and the second trend feature, and determining first seasonal convolution parameters based on the first activation function, the first seasonal feature, and the second seasonal feature; determining second trend convolution parameters based on a second activation function, the second trend feature, and the first trend feature, and determining second seasonal convolution parameters based on the second activation function, the second seasonal feature, and the first seasonal feature; obtaining the trend time series component output result based on the first trend convolution parameters and the second trend convolution parameters; and obtaining the seasonal time series component output result based on the first seasonal convolution parameters and the second seasonal convolution parameters.

[0149] Optionally, a sub-unit is established, including: constructing a preset dimension spatial feature of the seasonal time series component based on the first load point, the second load point, and preset time window parameters, wherein the preset dimension spatial feature is used to represent the spatial features in the preset dimension; and performing a linear transformation on the preset dimension spatial feature to obtain the time window load feature matrix.

[0150] Optionally, the lateral time window analyzer includes: a local multilayer perceptron, a first normalization layer, and a first feature extraction subunit, comprising: using the local multilayer perceptron to extract features from the time window load features at multiple time points to obtain first time window load features at multiple time points; using the first normalization layer to normalize the first time window load features at multiple time points to obtain second time window load features at multiple time points; and transposing the second time window load features at multiple time points to obtain lateral time window load features at multiple time points.

[0151] Optionally, the longitudinal time window analyzer includes: a global multilayer perceptron, a second normalization layer, and a second feature extraction subunit, which includes: using the global multilayer perceptron to extract features from the load features of the horizontal time window at multiple time points to obtain the load features of the third time window at multiple time points; using the second normalization layer to normalize the load features of the third time window at multiple time points to obtain the load features of the fourth time window at multiple time points; and transposing the load features of the fourth time window at multiple time points to obtain the comprehensive load features of the time window at multiple time points.

[0152] According to another aspect of the present invention, an electronic device is also provided, comprising: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods of various embodiments of the present invention during runtime.

[0153] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is executed, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of the present invention.

[0154] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of the present invention.

[0155] According to another aspect of the present invention, a computer program product is also provided, including a non-volatile computer-readable storage medium for storing a computer program that, when executed by a processor, implements the methods of various embodiments of the present invention.

[0156] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0157] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0158] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0159] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0160] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0161] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the power load of a power system, characterized in that, include: Historical power load data of the power system is obtained. The historical power load data includes multiple power load time series, wherein different power load time series are used to represent the time series corresponding to different variables affecting the power load. The multiple power load time series are input into the power load prediction model, and the power load prediction model is used to predict the multiple power load time series of the power system in the future time period to obtain predicted power load data, wherein the predicted power load data is used to represent the predicted power load data of the power system in the future time period. The power load forecasting model includes a time series decomposition module, a feature extraction module, and a feature enhancement module. Multiple power load time series are input into the power load forecasting model, which then predicts the power system's power load over a future time period to obtain target predicted power load data. This includes: decomposing the multiple power load time series using the time series decomposition module to obtain trend time series components and seasonal time series components, where the seasonal time series components are those exhibiting periodic fluctuations, and the trend time series components are those exhibiting long-term stable trends; extracting features from the seasonal time series components using the feature extraction module to obtain comprehensive time window load features; and enhancing the comprehensive time window load features and the trend time series components using the feature enhancement module to obtain the predicted power load data. The feature extraction module includes a time window embedding layer, a horizontal time window analyzer, and a vertical time window analyzer. The feature extraction module extracts features from the seasonal time series components to obtain comprehensive time window load features. This includes: using the time window embedding layer to establish a time window load feature matrix for a first load point and a second load point in the seasonal time series components, where the second load point represents a load point within an adjacent time window of the first load point, and the time window load feature matrix represents the time window load features at multiple times; using the horizontal time window analyzer to extract the time window load features at multiple times to obtain horizontal time window load features at multiple times; and using the vertical time window analyzer to extract the horizontal time window load features at multiple times to obtain comprehensive time window load features at multiple times.

2. The power load forecasting method for a power system according to claim 1, characterized in that, The predicted power load data is obtained by using a feature enhancement module to enhance the comprehensive time window load characteristics and the trend time series components, including: The feature enhancement module is used to enhance the comprehensive time window load features to obtain the seasonal time series component output results; The feature enhancement module is used to enhance the features of the trend time series component to obtain the output result of the trend time series component. The predicted power load data is determined based on the output results of the trend time series component and the output results of the seasonal time series component.

3. The power load forecasting method for a power system according to claim 2, characterized in that, The feature enhancement module includes a first convolutional layer and a second convolutional layer. The feature enhancement module is used to enhance the comprehensive time window load features to obtain the seasonal time series component output. The feature enhancement module is also used to enhance the trend time series component to obtain the trend time series component output, including: The first convolutional layer is used to extract features from the trend time series components and the comprehensive time window load features to obtain the first trend feature and the first seasonal feature. The second convolutional layer is used to extract features from the trend time series components and the comprehensive time window load features to obtain the second trend feature and the second seasonal feature, wherein the kernel size of the second convolutional layer is larger than the kernel size of the first convolutional layer; The first trend convolution parameters are determined based on the first activation function, the first trend feature, and the second trend feature; the first seasonal convolution parameters are determined based on the first activation function, the first seasonal feature, and the second seasonal feature. The second trend convolution parameters are determined based on the second activation function, the second trend feature, and the first trend feature; the second seasonal convolution parameters are determined based on the second activation function, the second seasonal feature, and the first seasonal feature. Based on the first trend convolution parameters and the second trend convolution parameters, the output result of the trend time series component is obtained; Based on the first seasonal convolution parameters and the second seasonal convolution parameters, the output result of the seasonal time series component is obtained.

4. The power load forecasting method for a power system according to claim 1, characterized in that, The time window embedding layer is used to establish a time window load feature matrix for the first load point and the second load point in the seasonal time series component, including: Based on the first load point, the second load point, and the preset time window parameters, a preset dimension spatial feature of the seasonal time series component is constructed, wherein the preset dimension spatial feature is used to represent the spatial features in the preset dimension; The preset dimensional spatial features are linearly transformed to obtain the time window load feature matrix.

5. The power load forecasting method for a power system according to claim 1, characterized in that, The lateral time window analyzer includes: a local multilayer perceptron and a first normalization layer. It analyzes the local correlation of the time window load characteristics at multiple time points to obtain the lateral time window load characteristics at multiple time points, including: The local multilayer perceptron is used to extract the load features of the time windows at the multiple time points to obtain the first time window load features at the multiple time points. The first time window load characteristics at multiple times are normalized using the first standardization layer to obtain the second time window load characteristics at multiple times. The load characteristics of the second time window at the multiple time points are transposed to obtain the load characteristics of the horizontal time window at the multiple time points.

6. The power load forecasting method for a power system according to claim 1, characterized in that, The longitudinal time window analyzer includes a global multilayer perceptron and a second normalization layer. It extracts the lateral time window load characteristics from the multiple time points to obtain the comprehensive time window load characteristics, including: The global multilayer perceptron is used to extract features from the horizontal time window load features at the multiple time points to obtain the third time window load features; The load characteristics of the third time window at the multiple time points are normalized using the second standardization layer to obtain the load characteristics of the fourth time window at the multiple time points. The load characteristics of the fourth time window at the multiple time points are transposed to obtain the comprehensive load characteristics of the multiple time points.

7. A power load forecasting device for a power system, characterized in that, include: The acquisition module is used to acquire historical power load data of the power system. The historical power load data includes multiple power load time series, wherein different power load time series are used to represent the time series corresponding to different variables affecting the power load. The prediction module is used to input the multiple power load time series into the power load prediction model, and use the power load prediction model to predict the multiple power load time series of the power system in the future time period to obtain predicted power load data, wherein the predicted power load data is used to represent the predicted power load data of the power system in the future time period. The power load prediction model in the prediction module includes a time series decomposition module, a feature extraction module, and a feature enhancement module. The prediction module further utilizes the time series decomposition module to decompose the multiple power load time series to obtain trend time series components and seasonal time series components. The seasonal time series components are those exhibiting periodic fluctuations in the multiple power load time series, and the trend time series components are those exhibiting long-term stable trends. The feature extraction module extracts features from the seasonal time series components to obtain comprehensive time window load features. The feature enhancement module enhances the comprehensive time window load features and the trend time series components to obtain the predicted power load data. The feature extraction module includes a time window embedding layer, a horizontal time window analyzer, and a vertical time window analyzer. The prediction module is further configured to use the time window embedding layer to establish a time window load feature matrix for the first load point and the second load point in the seasonal time series component, wherein the second load point is used to represent the load point within the adjacent time window of the first load point, and the time window load feature matrix is ​​used to represent the time window load features at multiple times; the horizontal time window analyzer is used to extract the time window load features at multiple times to obtain the horizontal time window load features at multiple times; and the vertical time window analyzer is used to extract the horizontal time window load features at multiple times to obtain the comprehensive time window load features at multiple times.

8. An electronic device, characterized in that, include: Memory, which stores executable programs; A processor for running the program, wherein the program, when running, performs the method according to any one of claims 1 to 6.

Images