Planning-oriented multi-scale peak load forecasting method based on attention network

Abstract

This application discloses a planning-oriented multi-scale peak load forecasting method based on attention networks, belonging to the field of attention network technology. The method includes: determining a target power grid; collecting historical window load data of the target power grid according to a preset time window; constructing multi-scale power grid window load characteristics by combining an attention mechanism; setting a planning-oriented error penalty factor; training and generating a peak load predictor based on the error penalty factor; processing the multi-scale power grid window load characteristics through the peak load predictor to obtain peak load prediction values; and generating a power grid equipment expansion plan based on the peak load prediction values. This method solves the technical problems of existing forecasting methods neglecting the actual economic impact of forecasting errors on power grid planning and lacking effective fusion and utilization of multi-time-scale load characteristics, resulting in low accuracy and reliability of forecasting results.

Description

Technical Field

[0001] This application relates to the field of attention network technology, specifically to a planning-oriented multi-scale peak load prediction method based on attention networks. Background Technology

[0002] With the continuous development of power systems and the increasing demand for electricity, accurate prediction of peak loads is crucial for the safe and stable operation of the power grid and the rational planning of equipment. However, traditional prediction methods focus only on prediction accuracy while neglecting the actual economic impact of prediction errors on power grid planning, and lack effective integration and utilization of load characteristics across multiple time scales. This results in low accuracy and reliability of prediction results, making them unable to adequately meet the actual needs of power grid planning. Summary of the Invention

[0003] This application provides a planning-oriented multi-scale peak load forecasting method based on attention networks, which solves the technical problems of existing forecasting methods ignoring the actual economic impact of forecasting errors on power grid planning and lacking effective fusion and utilization of load characteristics at multiple time scales, resulting in low accuracy and reliability of forecasting results.

[0004] The technical solution to the above-mentioned technical problems in this application is as follows:

[0005] In a first aspect, this application provides a planning-oriented multi-scale peak load prediction method based on attention networks, the method comprising:

[0006] Determine the target power grid, collect historical window load data of the target power grid according to a preset time window, and construct multi-scale power grid window load characteristics by combining an attention mechanism;

[0007] Set a planning-oriented error penalty factor, and train a peak load predictor based on the error penalty factor;

[0008] The peak load predictor processes the multi-scale power grid window load characteristics to obtain peak load prediction values, and generates power grid equipment expansion schemes based on the peak load prediction values.

[0009] Optionally, historical window load data of the target power grid are collected according to a preset time window, and multi-scale power grid window load characteristics are constructed by combining an attention mechanism, including:

[0010] Receive a preset time window uploaded by the user terminal, wherein the preset time window is greater than or equal to the minimum time window;

[0011] Collect historical window load data of the target power grid within the preset time window;

[0012] Multiple preset time scales are obtained, and the historical window load data is processed based on the multiple preset time scales to obtain load datasets with multiple scales.

[0013] Based on the attention mechanism, multiple attention feature extraction networks with multiple preset time scales are constructed. The multiple attention feature extraction networks are used to process the multiple scale load datasets respectively to obtain multiple scale load features.

[0014] The multi-scale load characteristics are fused to obtain the multi-scale power grid window load characteristics.

[0015] Optionally, multiple preset time scales are obtained, and the historical window load data is processed based on the multiple preset time scales to obtain multiple scale load datasets, including:

[0016] The plurality of preset time scales include a first preset time scale, a second preset time scale, and a third preset time scale;

[0017] The historical window load data is divided according to the first preset time scale, the second preset time scale, and the third preset time scale respectively to obtain a first-scale load dataset, a second-scale load dataset, and a third-scale load dataset;

[0018] The first-scale load dataset, the second-scale load dataset, and the third-scale load dataset are combined to obtain the multiple scale load datasets.

[0019] Optionally, the first preset time scale is an hourly time scale, the second preset time scale is a daily time scale, and the third preset time scale is a monthly time scale.

[0020] Optionally, multiple attention feature extraction networks with multiple preset time scales are constructed based on the attention mechanism, and the multiple scale load datasets are processed by the multiple attention feature extraction networks respectively to obtain multiple scale load features, including:

[0021] The attention mechanism includes multiple weight allocation strategies for multiple preset time scales, and each weight allocation strategy is the importance weight of data in different time periods within the corresponding preset time scale.

[0022] Based on each of the weight allocation strategies, attention feature extraction networks corresponding to preset time scales are constructed to obtain multiple attention feature extraction networks;

[0023] Extract the first scale load dataset from the multiple scale load datasets, and select the corresponding first attention feature extraction network from the multiple attention feature extraction networks;

[0024] The first attention feature extraction network is used to extract features from the first scale load dataset to generate first scale load features.

[0025] Following the method of generating the first-scale load features of the first-scale load dataset, the scale load features corresponding to the other scale load datasets are obtained, resulting in multiple scale load features.

[0026] Optionally, attention feature extraction networks corresponding to preset time scales are constructed based on each of the weight allocation strategies, resulting in multiple attention feature extraction networks, including:

[0027] Among the plurality of weight allocation strategies, a first weight allocation strategy is determined, and a corresponding first preset time scale is determined;

[0028] A first data fusion layer is constructed based on the first preset time scale. The first data fusion layer is used to fuse load data of the same time period in the first scale load data to obtain multiple time period fusion features.

[0029] A first weight allocation layer is constructed according to the first weight allocation strategy. The first weight allocation layer is used to allocate corresponding importance weights to the fusion features of each time period.

[0030] The first data fusion layer and the first weight allocation layer are combined to form the first attention feature extraction network;

[0031] Following the method used to form the first attention feature extraction network, attention feature extraction networks corresponding to other preset time scales are constructed to obtain multiple attention feature extraction networks.

[0032] Optionally, a planning-oriented error penalty factor is set, and a peak load predictor is generated based on the error penalty factor through training, including:

[0033] Based on the grid size of the target power grid, historical load data of similar power grids are collected;

[0034] The historical load data of the same type of power grid are processed according to the method of constructing the multi-scale power grid window load characteristics to obtain a sample multi-scale power grid window load feature set;

[0035] Based on the historical load data of the same type of power grid, the peak load of each sample multi-scale power grid window load feature in the sample multi-scale power grid window load feature set is labeled to obtain the sample peak load label value set.

[0036] Set planning-oriented error penalty factors, including a first penalty factor for adjusting load underestimation error and a second penalty factor for adjusting load overestimation error, wherein the first penalty factor is greater than the second penalty factor;

[0037] The peak load predictor is generated by training based on the sample multi-scale power grid window load feature set, the sample peak load label value set, and the error penalty factor.

[0038] Optionally, the peak load predictor is trained and generated based on the sample multi-scale power grid window load feature set, the sample peak load label value set, and the error penalty factor, including:

[0039] Construct a peak load predictor framework;

[0040] The sample multi-scale power grid window load feature set is used as the training input data, and the sample peak load label value set is used as the training target data.

[0041] The training input data is input into the peak load predictor framework to obtain the training prediction output value;

[0042] Calculate the prediction error between the training prediction output value and the training target data, and decompose the prediction error into load underestimation error and load overestimation error;

[0043] The load underestimation error is weighted using the first penalty factor, and the load overestimation error is weighted using the second penalty factor to obtain the weighted prediction error;

[0044] The network parameters of the peak load predictor framework are adjusted based on the weighted prediction error, and the training process is repeated until convergence is obtained to obtain the peak load predictor.

[0045] Optionally, the peak load predictor processes the multi-scale power grid window load characteristics to obtain a peak load prediction value, and generates a power grid equipment expansion plan based on the peak load prediction value, including:

[0046] The multi-scale power grid window load characteristics are input into the peak load predictor to obtain the peak load prediction value;

[0047] The expected equipment capacity requirement is determined based on the peak load forecast and the preset safety margin parameters;

[0048] Obtain the current equipment capacity information of the target power grid, compare and analyze the current equipment capacity information with the expected equipment capacity demand, and determine the equipment capacity gap;

[0049] Based on the aforementioned equipment capacity gap, an equipment expansion strategy is formulated, and the power grid equipment expansion plan is generated.

[0050] Optionally, determining the expected equipment capacity requirement based on the peak load forecast and preset safety margin parameters includes:

[0051] Obtain preset safety margin parameters, which include the upper limit of equipment load rate and capacity redundancy coefficient;

[0052] Calculate the basic equipment capacity requirement based on the peak load forecast and the upper limit of equipment load rate;

[0053] The capacity requirements of the basic equipment are adjusted according to the capacity redundancy coefficient to obtain the adjusted equipment capacity requirements, and the adjusted equipment capacity requirements are used as the expected equipment capacity requirements.

[0054] This application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0055] This application provides a planning-oriented multi-scale peak load forecasting method based on attention networks. First, it constructs a multi-temporal-scale data processing framework, simultaneously extracting hourly, daily, and monthly load features. An attention mechanism is used to identify the importance weights of data from different time periods, achieving intelligent fusion of cross-scale features. Second, a planning-oriented error penalty mechanism is introduced during model training, setting a higher penalty factor for load underestimation errors, giving the forecasting model a natural planning-friendly preference of "preferring moderate overestimation to avoiding underestimation." Finally, the forecast results are directly applied to the generation of power grid equipment expansion schemes, forming a complete closed loop from data acquisition to planning decision-making.

[0056] The above technical solutions provide strong guarantees for the safe and stable operation of the power grid, avoiding equipment overload damage due to load underestimation and reducing the probability of power outages. Simultaneously, they also prevent the waste of equipment resources caused by overestimating load, reducing power grid construction and operation costs. By effectively integrating and utilizing load characteristics across multiple time scales, the changing patterns of power grid load can be accurately captured, adapting to the electricity demand characteristics of different seasons and time periods, making power grid equipment expansion plans more scientific and rational, and better meeting the actual needs of power grid planning. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1This is a flowchart illustrating the planning-oriented multi-scale peak load prediction method based on attention networks provided in this application embodiment;

[0059] Figure 2 This is a schematic diagram of the process for generating power grid equipment expansion schemes in the planning-oriented multi-scale peak load prediction method based on attention networks provided in the embodiments of this application. Detailed Implementation

[0060] This application provides a planning-oriented multi-scale peak load forecasting method based on attention networks to address the technical problems of existing forecasting methods neglecting the actual economic impact of forecasting errors on power grid planning and lacking effective fusion and utilization of load characteristics across multiple time scales, resulting in low accuracy and reliability of forecasting results.

[0061] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0063] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0064] Example 1, as Figure 1 As shown, this application provides a planning-oriented multi-scale peak load prediction method based on attention networks, including:

[0065] S10: Determine the target power grid, collect historical window load data of the target power grid according to a preset time window, and construct multi-scale power grid window load characteristics by combining an attention mechanism;

[0066] In this embodiment, firstly, the target power grid is determined. When collecting historical load data of the target power grid according to a preset time window, the periodicity, seasonality, and trend of the power grid load are considered, and a preset time window is set. For example, for a power grid with obvious daily cycle characteristics, the preset time window can be relatively long to cover multiple complete daily cycles; for a power grid that is greatly affected by seasonal factors, a sufficiently long time window needs to be set to include data from different seasons.

[0067] Furthermore, when using historical window load data of the target power grid, an attention mechanism is incorporated to construct multi-scale power grid window load features, enabling the mining and integration of features from different scales. The attention mechanism can assign weights based on the importance of data from different time periods, allowing the model to focus more on time periods that have a significant impact on load forecasting.

[0068] Specifically, step S10 in the method includes:

[0069] Receive a preset time window uploaded by the user terminal, wherein the preset time window is greater than or equal to the minimum time window;

[0070] Collect historical window load data of the target power grid within the preset time window;

[0071] Multiple preset time scales are obtained, and the historical window load data is processed based on the multiple preset time scales to obtain load datasets with multiple scales.

[0072] Based on the attention mechanism, multiple attention feature extraction networks with multiple preset time scales are constructed. The multiple attention feature extraction networks are used to process the multiple scale load datasets respectively to obtain multiple scale load features.

[0073] The multi-scale load characteristics are fused to obtain the multi-scale power grid window load characteristics.

[0074] In this embodiment, firstly, when receiving the preset time window uploaded by the user terminal, it is ensured that the preset time window is greater than or equal to the minimum time window, which is one year, to ensure that the collected data has sufficient representativeness and reliability. Within the preset time window, historical load data of the target power grid is collected. For the collected historical load data of the target power grid, data encryption and backup methods are used to ensure data security and integrity.

[0075] Secondly, when acquiring multiple preset time scales, it is essential to fully consider the different cycles and characteristics of power grid load changes. For example, different time scales such as hourly, daily, weekly, and monthly can be set to analyze and process historical window load data from multiple perspectives. After processing the historical window load data based on the preset time scales, multiple scale load datasets are obtained. Different statistical methods and algorithms can be applied to process the data. For instance, for the hourly scale, statistical indicators such as average load, maximum load, and minimum load per hour can be calculated; while for the daily scale, information such as daily peak and off-peak electricity consumption periods and electricity consumption can be analyzed.

[0076] Furthermore, multiple attention feature extraction networks with preset time scales are constructed based on the attention mechanism, and the network structure and parameters are optimized. For example, convolutional neural networks from deep learning are used to construct the attention feature extraction networks. By continuously adjusting parameters such as the number of layers, nodes, and learning rate, the network performance and feature extraction capabilities are improved. When using multiple attention feature extraction networks to process load datasets at multiple scales, a cross-scale attention mechanism is adopted, allowing load features at different scales to influence and complement each other, thereby better capturing the complex changing patterns of power grid load.

[0077] Furthermore, the cross-scale attention mechanism refers to establishing connections between attention feature extraction networks at different time scales, allowing each network to reference information from other scales when extracting load features at its own scale. For example, when processing hourly load data, the network can adjust its level of attention to different hourly data based on load trends at daily or monthly scales.

[0078] Then, load characteristics at multiple scales are fused. For example, a weighted summation method is used to assign corresponding weights to load characteristics at different scales based on their importance to load forecasting. The weighted characteristics are then summed to obtain multi-scale power grid window load characteristics.

[0079] Through the above process of constructing multi-scale power grid window load characteristics, the variation pattern of power grid load at different time scales is fully explored, and the attention mechanism is used to highlight the data characteristics of important time periods, thereby achieving effective fusion of cross-scale characteristics.

[0080] This involves acquiring multiple preset time scales, processing the historical window load data based on these preset time scales, and obtaining load datasets at multiple scales, including:

[0081] The plurality of preset time scales include a first preset time scale, a second preset time scale, and a third preset time scale;

[0082] The historical window load data is divided according to the first preset time scale, the second preset time scale, and the third preset time scale respectively to obtain a first-scale load dataset, a second-scale load dataset, and a third-scale load dataset;

[0083] The first-scale load dataset, the second-scale load dataset, and the third-scale load dataset are combined to obtain the multiple scale load datasets.

[0084] In this embodiment, historical window load data is divided according to different time scale characteristics. The preset time scales include a first preset time scale, a second preset time scale, and a third preset time scale. The first preset time scale, the second preset time scale, and the third preset time scale represent different levels of refinement in dividing the preset time.

[0085] Then, the first-scale load dataset, the second-scale load dataset, and the third-scale load dataset are combined to obtain multiple-scale load datasets. These scaled load datasets comprehensively reflect the changing patterns of power grid load from multiple perspectives, providing a richer and more accurate data foundation for subsequent load forecasting and power grid planning. During the combination process, the accuracy and consistency of the data are ensured, and any potential outliers are processed and corrected to guarantee the quality of the multiple-scale load datasets.

[0086] Furthermore, the first preset time scale is an hourly time scale, the second preset time scale is a daily time scale, and the third preset time scale is a monthly time scale.

[0087] In this embodiment of the application, the first preset time scale is an hourly time scale, the second preset time scale is a daily time scale, and the third preset time scale is a monthly time scale.

[0088] Among them, the hourly time scale can accurately capture the fluctuations of the power grid load in a short period of time. For example, during different shifts in industrial production and peak and off-peak periods in commercial activities, the load will change rapidly and significantly. The hourly time scale data can clearly reflect the characteristics of the change. The daily time scale focuses on showing the daily periodicity of the power grid load, such as the difference in electricity consumption patterns on weekdays and weekends, as well as the peak and off-peak characteristics of electricity consumption at different times of the day. The monthly time scale is more conducive to grasping the seasonal and long-term trend changes of the power grid load. Due to the influence of factors such as climate, holidays, and production activities, electricity demand will show obvious fluctuations in different months. Monthly division helps to observe the seasonal changes of the power grid load.

[0089] For example, the load may differ significantly between working hours and rest periods on a weekday. When divided by day, the load data of each day is treated as a whole to obtain a second-scale load dataset. By analyzing the daily load data, the daily periodic changes in the power grid load can be discovered, and the load patterns on weekends and weekdays may be different.

[0090] Furthermore, historical window load data is divided into first-scale load dataset, second-scale load dataset, and third-scale load dataset according to the first preset time scale, the second preset time scale, and the third preset time scale. Then, multiple attention feature extraction networks are constructed to process the scale load datasets and obtain multiple scale load features.

[0091] Specifically, multiple attention feature extraction networks with multiple preset time scales are constructed based on the attention mechanism. These networks are then used to process the multiple scale load datasets to obtain multiple scale load features, including:

[0092] The attention mechanism includes multiple weight allocation strategies for multiple preset time scales, and each weight allocation strategy is the importance weight of data in different time periods within the corresponding preset time scale.

[0093] Based on each of the weight allocation strategies, attention feature extraction networks corresponding to preset time scales are constructed to obtain multiple attention feature extraction networks;

[0094] Extract the first scale load dataset from the multiple scale load datasets, and select the corresponding first attention feature extraction network from the multiple attention feature extraction networks;

[0095] The first attention feature extraction network is used to extract features from the first scale load dataset to generate first scale load features.

[0096] Following the method of generating the first-scale load features of the first-scale load dataset, the scale load features corresponding to the other scale load datasets are obtained, resulting in multiple scale load features.

[0097] In this embodiment, firstly, the attention mechanism employs multiple weight allocation strategies across multiple preset time scales to measure the importance of data from different time periods within the corresponding preset time scales, and then assigns weights to these preset time scales. For hourly time scales, since they reflect short-term load fluctuations, the weight allocation strategy may focus more on data from peak industrial production periods and busy commercial activity periods, as load changes during busy periods have a significant impact on overall load forecasting. For example, in industrially concentrated areas, factory start-up and shutdown times can cause significant fluctuations in hourly load; in this case, higher weights are assigned to data from critical time periods.

[0098] Secondly, attention feature extraction networks corresponding to preset time scales are constructed based on various weight allocation strategies, taking into account the characteristics of workload data at different time scales. When processing data of different scales through attention feature extraction networks, features can be extracted for each time period, and different features correspond to different importance weights. For hourly attention feature extraction networks, the structure and parameter settings should focus more on capturing rapidly changing workload features, and smaller convolutional kernels and shallower network layers can be used to improve sensitivity to short-term data changes. For monthly attention feature extraction networks, since monthly data reflects long-term trends and seasonal changes, larger convolutional kernels and deeper network layers can be used to better extract macroscopic features from the data.

[0099] Next, a first-scale load dataset is extracted from multiple scale load datasets, and a corresponding first-attention feature extraction network is selected. During feature extraction, the network's output is evaluated and adjusted. Cross-validation is used to divide the first-scale load dataset into training and validation sets. The parameters of the first-attention feature extraction network are optimized based on its performance on the validation set, ensuring that the generated first-scale load features accurately reflect the load characteristics at that time scale.

[0100] Then, a first-attention feature extraction network is used to extract features from the first-scale load dataset, generating first-scale load features. For example, local features in the first-scale load dataset are extracted through convolution operations, followed by feature dimensionality reduction through pooling operations. Finally, a fully connected layer maps the features to a low-dimensional space, resulting in a feature vector that represents the characteristics of the first-scale load data. The generated first-scale load features can be compared and analyzed with the actual load data to verify the effectiveness of the features. For example, it can be observed whether the features can accurately reflect the peak and trough periods of load at the hourly scale.

[0101] Finally, following the same method used to generate the first-scale load features of the first-scale load dataset, when obtaining the scale load features for other scale load datasets, such as the second-scale and third-scale load datasets, corresponding attention-based feature extraction networks are selected for feature extraction to ensure consistency and independence in the feature extraction process for each scale. Consistency is reflected in the fact that feature extraction is based on an attention-based weight allocation strategy, while independence requires that the feature extraction at each scale is not affected by other scales. For example, when generating daily-scale load features, only the daily-level weight allocation strategy and corresponding attention-based feature extraction network are used to avoid the influence of hourly or monthly data.

[0102] Among them, attention feature extraction networks corresponding to preset time scales are constructed based on each of the weight allocation strategies, resulting in multiple attention feature extraction networks, including:

[0103] Among the plurality of weight allocation strategies, a first weight allocation strategy is determined, and a corresponding first preset time scale is determined;

[0104] A first data fusion layer is constructed based on the first preset time scale. The first data fusion layer is used to fuse load data of the same time period in the first scale load data to obtain multiple time period fusion features.

[0105] A first weight allocation layer is constructed according to the first weight allocation strategy. The first weight allocation layer is used to allocate corresponding importance weights to the fusion features of each time period.

[0106] The first data fusion layer and the first weight allocation layer are combined to form the first attention feature extraction network;

[0107] Following the method used to form the first attention feature extraction network, attention feature extraction networks corresponding to other preset time scales are constructed to obtain multiple attention feature extraction networks.

[0108] In this embodiment, firstly, a first weight allocation strategy is determined among multiple weight allocation strategies, and a corresponding first preset time scale is also determined. When determining the first weight allocation strategy and the corresponding first preset time scale, the characteristics of power grid load changes and predicted demand at that time scale are considered. For example, for an hourly time scale, the first weight allocation strategy focuses on periods with a significant impact on load, such as industrial production and commercial operations.

[0109] Secondly, after determining the first preset time scale, a first data fusion layer is constructed. When fusing load data from the same time period within the first-scale load dataset, various fusion methods can be employed, such as weighted average fusion. Weighted average fusion can assign different weights based on the importance of different data points, making the fusion result more representative. Specifically, the first data fusion layer is used to fuse load data from the same time period within the first-scale load dataset, obtaining multiple time-period fusion features. These multiple time-period fusion features can more clearly demonstrate the overall load characteristics within the same time period at that time scale. For example, at an hourly time scale, fusing load data from different minutes within each hour yields the comprehensive load characteristics for each hour.

[0110] Next, a first weight allocation layer is constructed based on the first weight allocation strategy. This layer assigns importance weights to the fused features of each time period according to the strategy, resulting in the first-scale load features. When allocating weights, the specificity of each time period and its impact on overall load forecasting must be considered. Machine learning algorithms, such as linear regression, can be used to analyze historical data, identify the correlation between load data in different time periods and overall load changes, and thus determine appropriate weights. For example, for hourly time scales, if it is found that load changes in a certain industrial area during specific hours have a significant impact on the overall load for that day, then higher weights can be assigned to the fused features corresponding to those hours.

[0111] Then, the first data fusion layer and the first weight allocation layer are combined to form the first attention feature extraction network, which can extract representative features from the first-scale load dataset. During the combination process, the correctness of the data transfer and processing logic between the data fusion layer and the weight allocation layer is ensured to guarantee the normal operation of the network. Through the first attention feature extraction network, complex data information in the first-scale load dataset can be integrated and refined, highlighting data features of important time periods.

[0112] Finally, following the method used to form the first attention feature extraction network, attention feature extraction networks were constructed for other preset time scales, resulting in multiple attention feature extraction networks. For the daily time scale, the data fusion layer needs to consider the differences between weekdays and rest days when fusing load data from the same date; the weight allocation layer needs to consider the importance of peak and off-peak electricity consumption periods during the day when allocating weights. For the monthly time scale, the construction of the data fusion layer and the weight allocation layer should focus more on reflecting seasonal changes and long-term trends. During the construction process, each network was continuously debugged and optimized.

[0113] For example, for the second preset time scale, a second weight allocation strategy is first determined. Based on the characteristics of the daily load data, such as the differences between weekdays and rest days, and the peak and off-peak electricity consumption at different times of the day, the importance weights of data for different time periods are set. Next, a second data fusion layer is constructed to fuse load data from the same time period within the second-scale load dataset, such as load data from the same time period within a day. For example, a median fusion method can be used to reduce the impact of outliers, resulting in multiple time-period fusion features that reflect the overall load characteristics within the same daily time period. Then, a second weight allocation layer is constructed according to the second weight allocation strategy. Algorithms such as cluster analysis are used to study historical daily load data to identify load patterns on different dates and times, assigning reasonable importance weights to the fusion features of each time period. For example, the peak electricity consumption period on weekdays has a higher weight because it has a greater impact on overall daily load prediction. Finally, the second data fusion layer and the second weight allocation layer are combined to form a second attention feature extraction network, which can extract features reflecting the characteristics of the daily load from the second-scale load dataset.

[0114] S20: Set a planning-oriented error penalty factor, and train a peak load predictor based on the error penalty factor;

[0115] In this embodiment of the application, in order to more effectively control and adjust forecasting errors during the load forecasting process, a planning-oriented error penalty factor is set to meet the specific needs of power grid planning. The error penalty factor is set according to different planning objectives and requirements; for example, focusing on the accuracy of peak load forecasting, or setting different penalty levels for forecasting errors in different time periods.

[0116] The error penalty factor is an adjustable parameter that weights the prediction error. When the error between the predicted and actual values ​​is large, the error penalty factor increases the impact of this error during training, prompting the model to pay more attention to cases with larger errors, thereby improving the model's prediction accuracy on key metrics.

[0117] Furthermore, when setting the error penalty factor, various factors related to power grid planning need to be comprehensively considered. If the power grid planning focuses on ensuring power supply during peak hours, then a higher penalty factor can be set for prediction errors during peak hours. For example, during the summer peak electricity consumption period, the large-scale use of air conditioning and other cooling equipment leads to a sharp increase in power grid load. At this time, accurate prediction of peak load is crucial for the stable operation of the power grid. By increasing the error penalty factor during peak hours, the model will work harder to learn the load variation patterns during peak hours, reducing prediction errors.

[0118] Conversely, if power grid planning focuses more on load utilization efficiency during off-peak hours, a relatively high penalty factor can be set for prediction errors during these periods. During off-peak hours, the power grid load is relatively low, allowing some adjustable electrical equipment to operate at this time, thereby improving the overall utilization efficiency of the power grid.

[0119] Then, based on the error penalty factor, a peak load predictor is generated by training a neural network. During training, historical load data is input into the neural network. The model makes predictions based on the input data, calculates the error between the predicted and actual values, and weights the error using the error penalty factor. The parameters of the neural network are updated based on the weighted error, and the weights and parameters of the neural network are iteratively adjusted to enable the model to more accurately capture load change patterns in subsequent predictions.

[0120] Specifically, step S20 in the method includes:

[0121] Based on the grid size of the target power grid, historical load data of similar power grids are collected;

[0122] The historical load data of the same type of power grid are processed according to the method of constructing the multi-scale power grid window load characteristics to obtain a sample multi-scale power grid window load feature set;

[0123] Based on the historical load data of the same type of power grid, the peak load of each sample multi-scale power grid window load feature in the sample multi-scale power grid window load feature set is labeled to obtain the sample peak load label value set.

[0124] Set planning-oriented error penalty factors, including a first penalty factor for adjusting load underestimation error and a second penalty factor for adjusting load overestimation error, wherein the first penalty factor is greater than the second penalty factor;

[0125] The peak load predictor is generated by training based on the sample multi-scale power grid window load feature set, the sample peak load label value set, and the error penalty factor.

[0126] In this embodiment, firstly, historical load data of similar power grids are collected based on the grid size of the target power grid. Considering the differences in load characteristics between power grids of different sizes, when collecting historical load data of similar power grids, it is ensured that the power grid from which the data is collected is similar to the target power grid in terms of scale, electricity consumption structure, and regional characteristics. For example, if the target power grid is a medium-sized urban power grid, mainly for residential and commercial electricity consumption, then the collected historical load data of similar power grids should also come from urban power grids of similar scale and electricity consumption structure.

[0127] Secondly, historical load data of similar power grids were processed according to the method of constructing multi-scale power grid window load features to obtain a sample multi-scale power grid window load feature set. The processing followed the procedures and methods previously used to construct multi-scale power grid window load features. For data at different time scales, such as hourly, daily, and monthly, corresponding attention feature extraction networks were used for feature extraction and processing. When processing hourly data, an hourly-level attention feature extraction network was used, employing operations such as convolution, pooling, and fully connected layers to extract feature vectors that reflect hourly load characteristics.

[0128] Then, based on historical load data of similar power grids, peak load labels are applied to the load characteristics of each sample multi-scale power grid window in the sample multi-scale power grid window load feature set to obtain a sample peak load label value set. The labeling process involves identifying peak load points in the historical load data and mapping them to the corresponding sample multi-scale power grid window load characteristics. Peak load points can be determined by setting a load threshold; when the load data exceeds this threshold, it is marked as a peak load, and the corresponding feature information is recorded.

[0129] Next, planning-oriented error penalty factors are set, including a first penalty factor to adjust for load underestimation errors and a second penalty factor to adjust for load overestimation errors, with the first penalty factor being greater than the second. This is because in power grid planning, load underestimation may lead to insufficient power supply, affecting users' normal electricity consumption, and may even cause grid failures. While load overestimation may cause some resource waste, its harm is relatively smaller. Therefore, setting the first penalty factor to be greater than the second penalty factor emphasizes load underestimation during model training, reducing the occurrence of underestimation errors. For example, when the model predicts a load value lower than the actual value, a larger first penalty factor will allow the error to account for a larger proportion during training, prompting the model to adjust its parameters to improve the predicted load value.

[0130] Finally, a peak load predictor is trained based on a multi-scale power grid window load feature set, a sample peak load label set, and an error penalty factor. During training, optimization algorithms, such as stochastic gradient descent, are used to continuously adjust the model parameters. The sample multi-scale power grid window load features are input into the model, which outputs the predicted peak load. This predicted peak load is then compared with the sample peak load label values ​​to calculate the error, which is weighted using an error penalty factor. The model parameters are updated based on the weighted error. After multiple iterations of training, the model gradually converges, ultimately generating a predictor capable of accurately predicting peak loads.

[0131] The peak load predictor is generated by training based on the sample multi-scale power grid window load feature set, the sample peak load label value set, and the error penalty factor, including:

[0132] Construct a peak load predictor framework;

[0133] The sample multi-scale power grid window load feature set is used as the training input data, and the sample peak load label value set is used as the training target data.

[0134] The training input data is input into the peak load predictor framework to obtain the training prediction output value;

[0135] Calculate the prediction error between the training prediction output value and the training target data, and decompose the prediction error into load underestimation error and load overestimation error;

[0136] The load underestimation error is weighted using the first penalty factor, and the load overestimation error is weighted using the second penalty factor to obtain the weighted prediction error;

[0137] The network parameters of the peak load predictor framework are adjusted based on the weighted prediction error, and the training process is repeated until convergence is obtained to obtain the peak load predictor.

[0138] In this embodiment, a peak load predictor framework is first constructed, employing a neural network, which is selected and designed based on the characteristics of the load data and the prediction requirements. For example, for load data with time-series characteristics, neural networks are more suitable for capturing the time dependencies in the data.

[0139] Next, the sample multi-scale power grid window load feature set is used as the training input data, and the sample peak load label value set is used as the training target data. The sample multi-scale power grid window load feature set contains load feature information extracted from different time scales, reflecting the multifaceted characteristics of the power grid load. The sample peak load label value set is the actual peak load value from historical data, which is the target of model training. By inputting the data into the model, the model learns the mapping relationship between load features and peak load.

[0140] Secondly, after inputting the training input data into the peak load predictor framework, the training prediction output value is obtained. The training prediction output value is the model's prediction result of the peak load based on the input data.

[0141] For example, the peak load predictor is built and trained based on a neural network, and the specific steps are as follows:

[0142] First, data acquisition involves collecting a multi-scale power grid window load feature set and a sample peak load label value set. The input node of the peak load predictor framework is the sample peak load label value set.

[0143] Secondly, model construction involves building a convolutional neural network model, including convolutional layers, pooling layers, and fully connected layers. Convolutional layers extract features from the data, pooling layers reduce the size of the feature data, and fully connected layers convert load feature data into labels for predicted peak load. The input layer has a node count equal to the dimension of the input data. For example, if a multi-scale power grid window load feature set has 10 features, the input layer contains 10 nodes. One to three hidden layers are set, with the number of nodes in each layer adjusted experimentally (e.g., 64, 32, etc.). The ReLU activation function is used. The output layer has a node count equal to the predicted peak load; for example, the prediction time is represented by one node. The output layer generally does not use an activation function and directly outputs continuous values.

[0144] Next, the model is trained, and the predicted peak load is used as the output. The training framework is constructed using the Adam optimizer and the mean squared error loss function. The batch size is set to 32 and the total number of training epochs is 50. An early stopping mechanism (patience=5) is introduced. When the validation set loss does not decrease for 5 consecutive epochs, the training process is automatically terminated, resulting in a trained peak load predictor. This effectively avoids model overfitting while ensuring that the model reaches a convergent state.

[0145] Next, the prediction error between the training predicted output value and the training target data is calculated. The prediction error is an indicator of the model's prediction accuracy. The prediction error is decomposed into load underestimation error and load overestimation error. The load underestimation error is weighted using a first penalty factor, and the load overestimation error is weighted using a second penalty factor to obtain the weighted prediction error.

[0146] For example, in the historical data of a power grid, the training target data for a certain prediction was the actual peak load of 1000 MW, and the training prediction output value was 990 MW, resulting in a load underestimation error of 10 MW; the training target data for the second prediction was 890 MW, and the training prediction output value was 900 MW, resulting in a load overestimation error of 10 MW.

[0147] Assuming the first penalty factor is 0.2 and the second penalty factor is 0.1, then the weighted error of underestimating the load is 10 × 0.2 = 2, the weighted error of overestimating the load is 10 × 0.1 = 1, and the total weighted prediction error is 2 + 1 = 3.

[0148] The network parameters of the peak load predictor framework are adjusted based on weighted prediction error. In each training round, optimization algorithms such as gradient descent are used to update the weights and biases of each neuron in the neural network according to the magnitude and direction of the weighted prediction error. If the weighted prediction error is large, it indicates that the current parameter settings of the model are not reasonable and the parameters need to be adjusted significantly; if the weighted prediction error is small, the parameters can be fine-tuned.

[0149] Finally, the network parameters of the peak load predictor framework are adjusted based on the weighted prediction error. A gradient descent optimization algorithm is used to calculate the gradient of the network parameters based on the weighted prediction error, and then the parameters are updated in the opposite direction of the gradient. This training process is repeated until convergence. Convergence means that the model's performance no longer shows significant improvement, and the model parameters have reached a relatively stable state, ultimately resulting in a predictor that can accurately predict peak loads. During training, methods such as cross-validation are used to evaluate the model's performance and avoid overfitting and underfitting. Simultaneously, hyperparameters such as the error penalty factor are continuously adjusted to further optimize the model's performance.

[0150] By setting a planning-oriented error penalty factor and training a peak load predictor, the accuracy and reliability of peak load prediction can be improved.

[0151] S30: The peak load predictor processes the multi-scale power grid window load characteristics to obtain the peak load prediction value, and generates a power grid equipment expansion plan based on the peak load prediction value.

[0152] In this embodiment, after obtaining the trained peak load predictor, multi-scale grid window load characteristics are input into the predictor. These multi-scale grid window load characteristics contain various aspects of grid load information at different time scales. Based on the learned mapping relationship between load characteristics and peak load, the predictor processes and analyzes the grid window load characteristics to output the corresponding peak load prediction value. Based on the peak load prediction value, a grid equipment expansion plan is further generated.

[0153] Specifically, step S30 in the method includes:

[0154] The multi-scale power grid window load characteristics are input into the peak load predictor to obtain the peak load prediction value;

[0155] The expected equipment capacity requirement is determined based on the peak load forecast and the preset safety margin parameters;

[0156] Obtain the current equipment capacity information of the target power grid, compare and analyze the current equipment capacity information with the expected equipment capacity demand, and determine the equipment capacity gap;

[0157] Based on the aforementioned equipment capacity gap, an equipment expansion strategy is formulated, and the power grid equipment expansion plan is generated.

[0158] In this embodiment, firstly, since the multi-scale power grid window load characteristics are obtained through a large amount of historical data processing and analysis, reflecting the load characteristics of the power grid in different time dimensions, the multi-scale power grid window load characteristics are input into the peak load predictor. The predictor uses the mapping relationship between the load characteristics and the peak load learned in the previous training to analyze and calculate the input multi-scale power grid window load characteristics, thereby outputting the peak load prediction value.

[0159] Secondly, the expected equipment capacity demand is determined based on the peak load forecast and the preset safety margin parameter. The preset safety margin parameter is a coefficient set to cope with uncertainties and unforeseen circumstances in power grid operation. During power grid operation, some unforeseen factors may occur, such as extreme weather leading to a sudden increase in electricity demand, equipment failure, etc. The preset safety margin parameter can ensure that the power grid can still operate safely and stably under extreme conditions. For example, if the preset safety margin parameter is 0.1, and the peak load forecast is 1000 MW, then the expected equipment capacity demand is 1000 × (1 + 0.1) = 1100 MW.

[0160] Next, the current equipment capacity information of the target power grid is obtained through the power grid's equipment management system and related monitoring equipment. This current equipment capacity information includes transformer capacity, transmission line capacity, and distribution equipment capacity. The current equipment capacity information is then compared and analyzed with the expected equipment capacity demand to determine the equipment capacity gap. For example, if the current equipment capacity is 900 MW, and the expected equipment capacity demand is 1100 MW, then the equipment capacity gap is 1100 - 900 = 200 MW.

[0161] Finally, an equipment expansion strategy is formulated based on the capacity gap, generating a power grid equipment expansion plan. The expansion strategy needs to consider factors such as equipment procurement costs, installation cycles, and maintenance difficulty. Expansion methods include increasing transformer capacity, replacing transmission lines with thicker ones, and adding distribution equipment. For example, if the capacity gap is small, upgrading existing transformers can be considered; if the gap is large, new transformers or transmission lines may be needed. Simultaneously, the implementation time of the expansion plan must be considered, ideally during off-peak electricity demand periods to minimize impact on user power consumption.

[0162] The determination of expected equipment capacity requirements based on the peak load forecast and preset safety margin parameters includes:

[0163] Obtain preset safety margin parameters, which include the upper limit of equipment load rate and capacity redundancy coefficient;

[0164] Calculate the basic equipment capacity requirement based on the peak load forecast and the upper limit of equipment load rate;

[0165] The capacity requirements of the basic equipment are adjusted according to the capacity redundancy coefficient to obtain the adjusted equipment capacity requirements, and the adjusted equipment capacity requirements are used as the expected equipment capacity requirements.

[0166] In this embodiment, firstly, preset safety margin parameters are obtained. These parameters include an upper limit for equipment load rate and a capacity redundancy coefficient. The upper limit for equipment load rate is to ensure that the equipment operates in a safe and stable state, preventing damage due to overload. The capacity redundancy coefficient is to cope with possible emergencies, such as a sudden increase in electricity demand. For example, the upper limit for equipment load rate is set to 0.8, and the capacity redundancy coefficient is 0.1.

[0167] Next, the infrastructure capacity requirement is calculated based on the peak load forecast and the upper limit of equipment load rate. The calculation formula is: Infrastructure capacity requirement = Peak load forecast / Upper limit of equipment load rate. Assuming the peak load forecast is 800 MW and the upper limit of equipment load rate is 0.8, then the infrastructure capacity requirement = 800 / 0.8 = 1000 MW.

[0168] Then, the capacity requirement of the basic equipment is adjusted according to the capacity redundancy factor to obtain the adjusted equipment capacity requirement. The calculation formula is: Adjusted equipment capacity requirement = Basic equipment capacity requirement × (1 + Capacity redundancy factor). Based on the previously assumed data, the adjusted equipment capacity requirement = 1000 × (1 + 0.1) = 1100 MW.

[0169] Determining the expected equipment capacity requirements through the above methods allows for a reasonable consideration of various factors in power grid operation, providing an accurate basis for subsequent equipment expansion plans and ensuring that the power grid maintains a safe and stable operating state while meeting current and future electricity demands. Furthermore, in practice, it is necessary to dynamically adjust the preset safety margin parameters based on the actual conditions of the power grid to adapt to different operating environments and changing demands.

[0170] In summary, compared to existing technologies, this application improves the accuracy and reliability of peak load forecasting by setting a planning-oriented error penalty factor to weight underestimation and overestimation errors to varying degrees. This makes the model more focused on underestimation during training, effectively reducing underestimation errors. Regarding the formulation of power grid expansion plans, based on accurate peak load forecasts and preset safety margin parameters, the expected equipment capacity demand is determined. The equipment capacity gap is then identified through comparison with current equipment capacity information, leading to the formulation of expansion strategies. This fully considers uncertainties and unforeseen circumstances in power grid operation, such as extreme weather and equipment failures, ensuring the safe and stable operation of the power grid under various conditions. Furthermore, when determining expected equipment capacity demand, this application divides the preset safety margin parameters into an upper limit for equipment load rate and a capacity redundancy coefficient, considering both safe equipment operation and responses to unforeseen circumstances, providing a more accurate basis for power grid expansion. Simultaneously, it emphasizes the dynamic adjustment of preset safety margin parameters according to actual power grid conditions to adapt to different operating environments and demand changes, further improving the adaptability and stability of the power grid.

[0171] In summary, the embodiments of this application have at least the following technical effects:

[0172] This application provides a planning-oriented multi-scale peak load forecasting method based on attention networks. First, it constructs a multi-temporal-scale data processing framework, simultaneously extracting hourly, daily, and monthly load features. An attention mechanism is used to identify the importance weights of data from different time periods, achieving intelligent fusion of cross-scale features. Second, a planning-oriented error penalty mechanism is introduced during model training, setting a higher penalty factor for load underestimation errors, giving the forecasting model a natural planning-friendly preference of "preferring moderate overestimation to underestimation." Finally, the forecast results are directly applied to the generation of power grid equipment expansion plans, forming a complete closed loop from data acquisition to planning decision-making. This technical solution provides strong protection for the safe and stable operation of the power grid, avoiding equipment overload damage due to load underestimation and reducing the probability of power outages. Simultaneously, it avoids the waste of equipment resources caused by excessive load overestimation, reducing power grid construction and operation costs. By effectively integrating and utilizing load characteristics across multiple time scales, we can accurately capture the changing patterns of power grid load, adapt to the electricity demand characteristics of different seasons and time periods, make power grid equipment expansion plans more scientific and reasonable, better meet the actual needs of power grid planning, and promote the development of the power system towards a more intelligent, efficient, and reliable direction.

[0173] Among them, by processing the load characteristics of the multi-scale power grid window, a power grid equipment expansion scheme is generated, such as... Figure 2 As shown, it includes:

[0174] The multi-scale power grid window load characteristics are input into the peak load predictor to obtain the peak load prediction value;

[0175] The expected equipment capacity requirement is determined based on the peak load forecast and the preset safety margin parameters;

[0176] Obtain the current equipment capacity information of the target power grid, compare and analyze the current equipment capacity information with the expected equipment capacity demand, and determine the equipment capacity gap;

[0177] Based on the aforementioned equipment capacity gap, an equipment expansion strategy is formulated, and the power grid equipment expansion plan is generated.

[0178] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0179] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0180] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A planning-guided multi-scale peak load prediction method based on attention networks, characterized in that, The method includes: Determine the target power grid, collect historical window load data of the target power grid according to a preset time window, and construct multi-scale power grid window load characteristics by combining an attention mechanism; Set a planning-oriented error penalty factor, and train a peak load predictor based on the error penalty factor; The peak load predictor processes the multi-scale power grid window load characteristics to obtain peak load prediction values, and generates power grid equipment expansion schemes based on the peak load prediction values. The process includes setting a planning-oriented error penalty factor and training a peak load predictor based on the error penalty factor, comprising: Based on the grid size of the target power grid, historical load data of similar power grids are collected; The historical load data of the same type of power grid are processed according to the method of constructing the multi-scale power grid window load characteristics to obtain a sample multi-scale power grid window load feature set; Based on the historical load data of the same type of power grid, the peak load of each sample multi-scale power grid window load feature in the sample multi-scale power grid window load feature set is labeled to obtain the sample peak load label value set. Set planning-oriented error penalty factors, including a first penalty factor for adjusting load underestimation error and a second penalty factor for adjusting load overestimation error, wherein the first penalty factor is greater than the second penalty factor; The peak load predictor is generated by training based on the sample multi-scale power grid window load feature set, the sample peak load label value set, and the error penalty factor.

2. The method according to claim 1, characterized in that, Historical load data of the target power grid are collected according to a preset time window, and multi-scale power grid window load characteristics are constructed by combining an attention mechanism, including: Receive a preset time window uploaded by the user terminal, wherein the preset time window is greater than or equal to the minimum time window; Collect historical window load data of the target power grid within the preset time window; Multiple preset time scales are obtained, and the historical window load data is processed based on the multiple preset time scales to obtain load datasets with multiple scales. Based on the attention mechanism, multiple attention feature extraction networks with multiple preset time scales are constructed. The multiple attention feature extraction networks are used to process the multiple scale load datasets respectively to obtain multiple scale load features. The multi-scale load characteristics are fused to obtain the multi-scale power grid window load characteristics.

3. The method according to claim 2, characterized in that, Multiple preset time scales are obtained, and the historical window load data is processed based on the multiple preset time scales to obtain multiple scale load datasets, including: The plurality of preset time scales include a first preset time scale, a second preset time scale, and a third preset time scale; The historical window load data is divided according to the first preset time scale, the second preset time scale, and the third preset time scale respectively to obtain a first-scale load dataset, a second-scale load dataset, and a third-scale load dataset; The first-scale load dataset, the second-scale load dataset, and the third-scale load dataset are combined to obtain the multiple scale load datasets.

4. The method according to claim 3, characterized in that, The first preset time scale is an hourly time scale, the second preset time scale is a daily time scale, and the third preset time scale is a monthly time scale.

5. The method according to claim 3, characterized in that, Based on the aforementioned attention mechanism, multiple attention feature extraction networks with multiple preset time scales are constructed. These networks are then used to process the multiple scale load datasets to obtain multiple scale load features, including: The attention mechanism includes multiple weight allocation strategies for multiple preset time scales, and each weight allocation strategy is the importance weight of data in different time periods within the corresponding preset time scale. Based on each of the weight allocation strategies, attention feature extraction networks corresponding to preset time scales are constructed to obtain multiple attention feature extraction networks; Extract the first scale load dataset from the multiple scale load datasets, and select the corresponding first attention feature extraction network from the multiple attention feature extraction networks; The first attention feature extraction network is used to extract features from the first scale load dataset to generate first scale load features. Following the method of generating the first-scale load features of the first-scale load dataset, the scale load features corresponding to the other scale load datasets are obtained, resulting in multiple scale load features.

6. The method according to claim 5, characterized in that, Based on the weight allocation strategies described above, attention feature extraction networks corresponding to preset time scales are constructed respectively, resulting in multiple attention feature extraction networks, including: Among the plurality of weight allocation strategies, a first weight allocation strategy is determined, and a corresponding first preset time scale is determined; A first data fusion layer is constructed based on the first preset time scale. The first data fusion layer is used to fuse load data of the same time period in the first scale load data to obtain multiple time period fusion features. A first weight allocation layer is constructed according to the first weight allocation strategy. The first weight allocation layer is used to allocate corresponding importance weights to the fusion features of each time period. The first data fusion layer and the first weight allocation layer are combined to form the first attention feature extraction network; Following the method used to form the first attention feature extraction network, attention feature extraction networks corresponding to other preset time scales are constructed to obtain multiple attention feature extraction networks.

7. The method according to claim 1, characterized in that, The peak load predictor is trained and generated based on the sample multi-scale power grid window load feature set, the sample peak load label value set, and the error penalty factor, including: Construct a peak load predictor framework; The sample multi-scale power grid window load feature set is used as the training input data, and the sample peak load label value set is used as the training target data. The training input data is input into the peak load predictor framework to obtain the training prediction output value; Calculate the prediction error between the training prediction output value and the training target data, and decompose the prediction error into load underestimation error and load overestimation error; The load underestimation error is weighted using the first penalty factor, and the load overestimation error is weighted using the second penalty factor to obtain the weighted prediction error; The network parameters of the peak load predictor framework are adjusted based on the weighted prediction error, and the training process is repeated until convergence is obtained to obtain the peak load predictor.

8. The method according to claim 1, characterized in that, The peak load predictor processes the multi-scale power grid window load characteristics to obtain peak load prediction values, and generates a power grid equipment expansion plan based on the peak load prediction values, including: The multi-scale power grid window load characteristics are input into the peak load predictor to obtain the peak load prediction value; The expected equipment capacity requirement is determined based on the peak load forecast and the preset safety margin parameters; Obtain the current equipment capacity information of the target power grid, compare and analyze the current equipment capacity information with the expected equipment capacity demand, and determine the equipment capacity gap; Based on the aforementioned equipment capacity gap, an equipment expansion strategy is formulated, and the power grid equipment expansion plan is generated.

9. The method according to claim 8, characterized in that, Determining the expected equipment capacity requirement based on the peak load forecast and preset safety margin parameters includes: Obtain preset safety margin parameters, which include the upper limit of equipment load rate and capacity redundancy coefficient; Calculate the basic equipment capacity requirement based on the peak load forecast and the upper limit of equipment load rate; The capacity requirements of the basic equipment are adjusted according to the capacity redundancy coefficient to obtain the adjusted equipment capacity requirements, and the adjusted equipment capacity requirements are used as the expected equipment capacity requirements.

Images