Power source load multi-temporal-spatial-scale generality prediction method and related device

Abstract

The invention belongs to the technical field of power source load prediction, and discloses a power source load multi-temporal-spatial-scale generality prediction method and a related device. The power source load multi-temporal-spatial-scale generality prediction method comprises the following steps: based on a selected application scene and a prediction task, selecting a pre-training model from a pre-constructed model library; carrying out transfer learning by taking the selected pre-training model as a basic model and taking the selected application scene and the annotation data corresponding to the prediction task as a learning sample to obtain a fine-tuned power source load prediction model; and performing prediction by using the obtained power source load prediction model to obtain a prediction result. According to the technical scheme disclosed by the invention, a unified, dynamic and efficient power source load prediction model is established by fusing multi-source heterogeneous data, and the prediction accuracy, reliability, universality, flexibility and the like are greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power source and load forecasting, and particularly relates to a common forecasting method for power source and load with multiple spatio-temporal scales and related devices. Background Art

[0002] With the continuous improvement of the penetration rate of new energy, power system dispatching and forecasting face more challenges; specifically, in the process of spatio-temporal forecasting of power sources and loads, not only the volatility of short-term loads needs to be considered, but also comprehensive analysis needs to be carried out for multiple factors such as distributed energy, meteorological conditions, and market transactions. The existing forecasting schemes have the problem of lacking a unified framework and cannot achieve accurate forecasting across scales and dimensions.

[0003] Existing traditional power source and load forecasting models are mostly based on a single time scale, and it is difficult to adapt to the increasingly complex scenario requirements in modern power grids. There is a lack of an effective coordination mechanism between intraday load forecasting, short-term load forecasting, and ultra-long-term load forecasting, and a unified and complete forecasting system cannot be formed, resulting in low reliability of forecasting results and affecting the accuracy of power grid dispatching decisions; in addition, with the wide application of artificial intelligence and big data technologies, the requirements of power grid dispatching for real-time performance and accuracy are getting higher and higher, but the existing technical solutions still have deficiencies in data fusion, model updating, etc., and it is difficult to effectively meet the above requirements.

[0004] In summary, in view of the above-mentioned situations of power marketization, distributed energy access, and increased complexity of meteorological data, the existing technical solutions still have the following main problems: (1) Most of the existing power system forecasting methods are limited to a single data source or a single time scale, lacking effective fusion of multi-spatio-temporal and multi-dimensional data, resulting in insufficient accuracy and reliability of forecasting; (2) Although the forecasting methods based on artificial intelligence have significantly improved in accuracy, there is still a lack of a unified forecasting framework, making it difficult to perform effective model switching and fusion in different application scenarios, and the universality and flexibility of the models are poor. Summary of the Invention

[0005] The purpose of the present invention is to provide a common forecasting method for power source and load with multiple spatio-temporal scales and related devices to solve one or more of the above-mentioned technical problems. The technical solution disclosed by the present invention builds a unified, dynamic, and efficient power source and load forecasting model by fusing multi-source heterogeneous data, and performs forecasting based on the constructed power source and load forecasting model, which has greatly improved in terms of forecasting accuracy, reliability, universality, and flexibility.

[0006] To achieve the above object, the present invention adopts the following technical solutions: In the first aspect of the present invention, a common forecasting method for power source and load with multiple spatio-temporal scales is provided, including the following steps: Select a pre-trained model from a pre-constructed model library based on the selected application scenario and prediction task; Use the selected pre-trained model as the base model, and use the labeled data corresponding to the selected application scenario and prediction task as learning samples for transfer learning to obtain a fine-tuned power source and load prediction model; among them, when performing transfer learning, adjust the model parameters of the pre-trained model based on the source and load conditions and prediction result errors of the labeled data; Based on the selected application scenario and prediction task, use the obtained power source and load prediction model to make a prediction and obtain a prediction result; Among them, the pre-trained models stored in the pre-constructed model library are all power source and load pre-trained models obtained by training based on existing large-scale historical power load data, meteorological data, and distributed energy data.

[0007] A further improvement of the technical solution of the present invention lies in, The pre-trained model includes a data processing module, a gating module, and a multi-temporal and spatial scale prediction module; among them, The data processing module is used to map the input data into an original high-dimensional feature vector; The gating module is used to filter redundant features in the original high-dimensional feature vector to obtain a processed high-dimensional feature vector; The multi-temporal and spatial scale prediction module includes a variable selection network, a static covariate encoder, a temporal processing module, and a temporal self-attention decoder; among them, the variable selection network is used to input the processed high-dimensional feature vector, and based on the attention mechanism, select the input features most relevant to the current prediction task at each time step, and output the screened key features and their weights; the static covariate encoder is used to input the screened key features and their weights, and use the local perception and pooling operations of the convolutional neural network to construct multi-scale spatio-temporal dependence features, and output a context vector that fuses multi-scale spatio-temporal dependence; the temporal processing module is used to input the context vector that fuses multi-scale spatio-temporal dependence, perform local information processing and output temporal data that retains local temporal dependence; the temporal self-attention decoder is used to input the temporal data that retains local temporal dependence, capture the correlation between different positions in the sequence through the temporal self-attention mechanism, and combine local and global temporal information to output a prediction result.

[0008] A further improvement of the technical solution of the present invention lies in, In the data processing module, the step of mapping the input data into an original high-dimensional feature vector includes: Obtain the original data; among them, the original data includes power load data, meteorological data, distributed energy generation data, and preset time features; Preprocess the original data to obtain the preprocessed data. Among them, the steps of preprocessing include: filling in the missing parts in the power load data and meteorological data, and unifying the power load data, meteorological data, and distributed energy generation data to the same time node. Perform mapping processing based on the preprocessed data to obtain the embedded original high-dimensional feature vector.

[0009] A further improvement of the technical solution of the present invention lies in The power load data includes historical power load data of multiple time scales divided according to the collection frequency. The meteorological data includes one or more of temperature, humidity, wind speed, light, and rainfall, and is collected at time scales of minutes, hours, and days. The preset time features include one or more of season, holiday, and timestamp.

[0010] A further improvement of the technical solution of the present invention lies in The time self-attention decoder uses multi-head self-attention, and each head focuses on different preset time patterns.

[0011] A further improvement of the technical solution of the present invention lies in that in the step of obtaining the prediction result by using the obtained power source and load prediction model based on the selected application scenario and prediction task, the obtained prediction result is applicable to wind power prediction, photovoltaic prediction, bearing capacity assessment, heavy overload prediction, or intelligent dispatching decision-making.

[0012] In the second aspect of the present invention, a power source and load multi-time and space scale common prediction system is provided, including: A model selection module for selecting a pre-trained model from a pre-constructed model library based on the selected application scenario and prediction task. A model fine-tuning module for using the selected pre-trained model as the basic model and using the labeled data corresponding to the selected application scenario and prediction task as learning samples for transfer learning to obtain a fine-tuned power source and load prediction model. Among them, when performing transfer learning, adjust the model parameters of the pre-trained model based on the source and load situation and prediction result error of the labeled data. A model prediction module for obtaining a prediction result applicable to wind power prediction, photovoltaic prediction, bearing capacity assessment, heavy overload prediction, or intelligent dispatching decision-making by using the obtained power source and load prediction model based on the selected application scenario and prediction task. Among them, the pre-trained models stored in the pre-constructed model library are all power source and load pre-trained models obtained by training based on existing large-scale historical power load data, meteorological data, and distributed energy data.

[0013] A further improvement of the technical solution of the present invention lies in that The pre-trained model includes a data processing module, a gating module, and a multi-temporal and spatial scale prediction module; wherein, The data processing module is used to map the input data into an original high-dimensional feature vector; The gating module is used to filter redundant features in the original high-dimensional feature vector to obtain a processed high-dimensional feature vector; The multi-temporal and spatial scale prediction module includes a variable selection network, a static covariate encoder, a temporal processing module, and a temporal self-attention decoder; wherein, the variable selection network is used to input the processed high-dimensional feature vector, and based on the attention mechanism, select the input features most relevant to the current prediction task at each time step, and output the screened key features and their weights; the static covariate encoder is used to input the screened key features and their weights, and use the local perception and pooling operations of the convolutional neural network to construct multi-scale spatio-temporal dependence features, and output a context vector that fuses multi-scale spatio-temporal dependence; the temporal processing module is used to input the context vector that fuses multi-scale spatio-temporal dependence, perform local information processing and output temporal data that retains local temporal dependence relationships; the temporal self-attention decoder is used to input the temporal data that retains local temporal dependence relationships, capture the correlations between different positions in the sequence through the temporal self-attention mechanism, and combine local and global temporal information to output a prediction result.

[0014] A further improvement of the technical solution of the present invention lies in that In the data processing module, the step of mapping the input data into an original high-dimensional feature vector includes: Obtain original data; wherein, the original data includes power load data, meteorological data, distributed energy generation data, and preset time features; Preprocess the original data to obtain preprocessed data; wherein, the steps of preprocessing include: filling in the missing parts of the power load data and meteorological data, and unifying the power load data, meteorological data, and distributed energy generation data to the same time node; Perform mapping processing based on the preprocessed data to obtain the embedded original high-dimensional feature vector.

[0015] A further improvement of the technical solution of the present invention lies in that The power load data includes historical power load data of multiple time scales divided according to the acquisition frequency; The meteorological data includes one or more of temperature, humidity, wind speed, light, and rainfall, and is collected at time scales of minutes, hours, and days; The preset time features include one or more of season, holiday, and timestamp.

[0016] A further improvement of the technical solution of the present invention lies in that the time self-attention decoder adopts multi-head self-attention, and each head focuses on different preset time patterns.

[0017] A further improvement of the technical solution of the present invention lies in that in the step of obtaining a prediction result by using the obtained power source and load prediction model based on the selected application scenario and the prediction task, the obtained prediction result is applicable to wind power prediction, photovoltaic prediction, bearing capacity assessment, heavy overload prediction or intelligent scheduling decision-making.

[0018] In a third aspect of the present invention, there is provided an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, it implements the power source and load multi-temporal and spatial scale common prediction method according to any one of the first aspects of the present invention.

[0019] In a fourth aspect of the present invention, there is provided a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the power source and load multi-temporal and spatial scale common prediction method according to any one of the first aspects of the present invention.

[0020] Compared with the prior art, the present invention has the following beneficial effects: The common prediction method for power source and load in multiple spatio-temporal scales provided by the present invention establishes a unified, dynamic and efficient power source and load prediction model by integrating multi-source heterogeneous data. Based on the constructed power source and load prediction model, the prediction has been greatly improved in terms of prediction accuracy, reliability, universality and flexibility. Specifically, the pre-trained models stored in the pre-constructed model library are trained based on existing large-scale historical power load data, meteorological data and distributed energy data. Therefore, the present invention constructs a pre-trained model for multi-source data fusion. Through multi-source data fusion, the information of different data sources and different time scales can be fully utilized, improving the accuracy and reliability of the prediction. Based on the pre-constructed model library containing multiple pre-trained models, for different application scenarios and prediction tasks, a suitable pre-trained model is selected from the model library, and then the selected pre-trained model is used as the basic model, and the labeled data corresponding to the selected application scenario and prediction task is used as the learning sample for transfer learning; during the transfer learning process, the model parameters of the pre-trained model are adjusted based on the source load situation and prediction result error of a small amount of labeled data, so that the fine-tuned model can adapt to the specific application scenario. The above technical means of the present invention, through constructing a model library and a transfer learning mechanism, and adopting a unified model architecture, enable the model to effectively switch and fuse in different application scenarios, improving the universality and flexibility of the model, and being able to better adapt to the situation of power marketization, distributed energy access and increasing complexity of meteorological data.

[0021] In the preferred technical solution of the present invention, the pre-trained model includes a data processing module, a gating module, and a multi-temporal and spatial scale prediction module; the gating module filters redundant features in the original high-dimensional feature vector to obtain a processed high-dimensional feature vector; in the multi-temporal and spatial scale prediction module, the variable selection network selects the input features most relevant to the current prediction task at each time step based on the attention mechanism, the static covariate encoder constructs multi-scale spatio-temporal dependence features using the local perception and pooling operations of the convolutional neural network, the temporal processing module performs local information processing and outputs temporal data retaining local temporal dependence relationships, and the temporal self-attention decoder captures the correlation between different positions in the sequence through the temporal self-attention mechanism and outputs the prediction result, thus realizing the effective fusion and processing of multi-temporal, multi-dimensional data. Summarily, the core technical means of the technical solution of the present invention include: constructing a pre-trained model trained based on multi-source data, fusing and processing multi-source, multi-temporal and spatial data through the data processing module, the gating module, and the multi-temporal and spatial scale prediction module; establishing a model library containing multiple pre-trained models, adopting a transfer learning mechanism, and fine-tuning the pre-trained models according to different application scenarios and prediction tasks; designing a unified model architecture with a data processing module, a gating module, and a multi-temporal and spatial scale prediction module to provide a basis for model switching and fusion. The core technical effects of the technical solution of the present invention include: through multi-source data fusion and multi-module collaborative processing, making full use of the information of different data sources and different time scales, effectively solving the deficiencies of existing methods in multi-temporal, multi-dimensional data processing, and improving the accuracy and reliability of prediction; the model library and the transfer learning mechanism enable the model to adapt to different application scenarios, and the unified model architecture provides convenience for model switching and fusion, improving the universality and flexibility of the model, and being able to better cope with the complex changes of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art; obviously, the drawings in the following description are some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0023] Figure 1 It is a schematic flowchart of a method for predicting the commonality of power source and load at multiple temporal and spatial scales in an embodiment of the present invention; Figure 2 It is a schematic flowchart of predicting the commonality of power source and load at multiple temporal and spatial scales in a specific embodiment of the present invention; Figure 3 It is a schematic diagram of the data framework for predicting the commonality of power source and load at multiple temporal and spatial scales in a specific embodiment of the present invention; Figure 4It is a schematic diagram of the overall architecture of the common prediction of power source and load at multiple spatial and temporal scales in a specific embodiment of the present invention; Figure 5 It is a schematic diagram of the hardware support of the power source and load multi-spatial and temporal scale prediction system in a specific embodiment of the present invention; Figure 6 It is a schematic diagram of a power source and load multi-spatial and temporal scale common prediction system in an embodiment of the present invention. Specific embodiments

[0024] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention; obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments.

[0025] Based on the technical solutions disclosed in the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0026] Please refer to Figure 1 , a power source and load multi-spatial and temporal scale common prediction method provided by an embodiment of the present invention includes the following steps: Step 1, select a pre-trained model from a pre-constructed model library based on the selected application scenario and prediction task. Specifically, the selection of the application scenario and prediction task should be defined and classified according to the spatio-temporal characteristics of the power load. The application scenarios include power source and load prediction tasks at different time scales such as ultra-short-term prediction, short-term prediction, medium- and long-term prediction, and at different spatial scales such as regional level, urban level, and sub-district level. For ultra-short-term prediction, it is mainly used for real-time monitoring and emergency response, such as heavy overload analysis; short-term prediction is used for tasks such as dispatching plans; while long-term prediction is used for planning decisions, such as seasonal load fluctuation prediction. According to these scenarios, select a pre-trained model similar to the prediction task. Usually, the selection criteria can be based on the similarity of the application scenario and the consistency of the prediction task.

[0027] Step 2: Using the pre-trained model selected in Step 1 as the base model, and using the labeled data corresponding to the selected application scenario and prediction task as the learning samples for transfer learning to obtain a fine-tuned power source-load prediction model. During the transfer learning process, first analyze the spatio-temporal characteristics of the source-load based on the labeled data, and extract the load fluctuation patterns, seasonal variations, and the impact of special events on the load, etc. Subsequently, adjust the parameters of the pre-trained model through an error feedback mechanism. During the transfer learning process, automatically adjust the model parameters of the pre-trained model based on the source-load situation and prediction result error of the labeled data. Specifically, when the error between the prediction result and the real data is large, the model will dynamically adjust the learning rate and correct the weights of the model according to the error feedback. In addition, the system will also identify the time periods or spatial regions with large prediction errors through an adaptive error adjustment mechanism and make targeted corrections. Through this transfer learning method, the pre-trained model can fully adapt to the new power source-load prediction task and achieve a high prediction accuracy through automatic adjustment and optimization.

[0028] Step 3: Based on the application scenario and prediction task selected in Step 1, use the power source-load prediction model obtained in Step 2 for prediction to obtain prediction results applicable to wind power prediction, photovoltaic prediction, carrying capacity assessment, heavy overload prediction, or intelligent dispatching decision-making; Among them, the model library in Step 1 stores pre-trained models trained based on existing large-scale historical power load data, meteorological data, and distributed energy data.

[0029] The technical solution of the embodiment of the present invention constructs a cross-space-time and cross-scale power source-load prediction model by using the artificial intelligence methods of pre-trained models and transfer learning. The cross-space-time and cross-scale prediction solution provided by the present invention is not only applicable to the load prediction of traditional power grids, but also can handle complex scenarios after the access of distributed energy, and can accurately predict the power generation of distributed energy and its impact on the overall load to help the power grid dispatching center make optimized decisions; in addition, the technical solution of the embodiment of the present invention can also be applied to the power market trading platform to optimize the market trading strategy and improve the utilization rate of power resources by predicting electricity prices, load demands, and new energy output.

[0030] In a specific embodiment of the present invention, the pre-trained model includes: a data processing module, a gating module, and a multi-space-time scale prediction module; among them, The data processing module is used to map the input data into an original high-dimensional feature vector; The gating module is used to filter redundant features in the original high-dimensional feature vector to obtain a processed high-dimensional feature vector; The multi - spatio - temporal scale prediction module includes a variable selection network, a static covariate encoder, a time - series processing module, and a time self - attention decoder. Among them, the variable selection network is used to input the processed high - dimensional feature vector, automatically select the input variables most relevant to the current prediction task at each time step based on the attention mechanism, and output the selected key features and their weights. The static covariate encoder is used to input the selected key features and their weights, construct multi - scale spatio - temporal dependence features, and output a context vector that fuses multi - scale spatio - temporal dependence. The time - series processing module is used to input the context vector that fuses multi - scale spatio - temporal dependence, process local information in a sequence - to - sequence manner, and obtain time - series data that retains local time - series dependence relationships. The time self - attention decoder is used to input the time - series data that retains local time - series dependence relationships, automatically capture the correlation between different positions in the input sequence through the time self - attention mechanism, and output the prediction result.

[0031] In the preferred technical solution of the embodiment of the present invention, by obtaining meteorological data and distributed energy production data, a pre - training module for multi - source data fusion is constructed, enhancing the accuracy and reliability of prediction. In addition, the static variable encoder and the attention mechanism are used to effectively capture complex non - linear relationships at different time scales and spatial dimensions, improving the accuracy and applicability of prediction. Furthermore, by using transfer learning and model fine - tuning techniques, load forecasting for intraday, short - term, and long - term is realized, effectively integrating short - timeliness and long - timeliness requirements.

[0032] In the technical solution of the specific embodiment, in the grid day - ahead scheduling scenario: the input is future 24 - hour meteorological forecast data (temperature, wind speed) and historical load curves for the same period. The processing process is to call the short - term load forecasting model to generate sub - hourly load forecasts. The output is the unit commitment plan and the tie - line power plan.

[0033] In the technical solution of the specific embodiment, in the demand response management scenario, the input is real - time electricity price signals and user baseline loads. The processing process is to predict the user response potential and optimize the adjustable load scheduling. The output is the demand response strategy and the user incentive plan.

[0034] Please refer to Figures 2 to 5 , in the technical solution disclosed in the embodiment of the present invention, by integrating prediction models of different time scales and fusing multi - source heterogeneous data, a unified, dynamic, and efficient power source - load prediction model is established.

[0035] The embodiment of the present invention designs the overall process of the multi - spatio - temporal scale common prediction service for power source - load, constructs the technical architecture of the system, realizes cross - space - time and cross - scale source - load prediction, and the exemplary process of the prediction service is as Figure 2 shown.

[0036] The data collection and processing process is a key step in building a multi - spatio - temporal scale prediction system for power source and load. This process integrates various relevant data in the operation of the power system, including multi - source data such as historical power load, meteorological data, and distributed energy data. The original data collected is subjected to a series of pre - processing operations, providing high - quality inputs for model training and prediction.

[0037] In the specific exemplary technical solution of the embodiment of the present invention, power load data is obtained from the power grid operation monitoring system. The collection frequency during the data acquisition process can be set according to the prediction requirements, and can be specifically divided into multiple time scales such as intraday load, short - term load, and ultra - short - term load. Among them, intraday load data is collected at relatively short time intervals (such as 15 minutes or 30 minutes) to capture the rapid changes in power load within a day, which is crucial for analyzing the load characteristics during peak and off - peak hours of electricity consumption. Short - term load data is usually collected in hours to reflect the load trend changes within several days to several weeks, helping grid dispatching personnel formulate power generation plans for a week in advance. Ultra - short - term load data has a higher collection frequency (such as 5 minutes or even shorter), mainly used for real - time monitoring of the dynamic balance of the power system and promptly detecting sudden load fluctuations. In addition, meteorological data including parameters such as temperature, humidity, wind speed, sunlight, and rainfall is obtained in real - time through the enterprise - level meteorological data service center. Meteorological data directly affects the power generation capacity of photovoltaic and wind power. Meteorological data is collected at minute - level, hour - level, and daily - level time scales to ensure that the prediction system can quickly respond to short - term climate changes.

[0038] In a specific embodiment of the present invention, to ensure the accuracy and consistency of the model input data, all the collected original data needs to be pre - processed first, and the pre - processed data is then applied to the model. The specific exemplary pre - processing steps include: First, clear the missing values, outliers, and noise data in the data. Among them, for the missing parts in meteorological data and load data, interpolation methods or the average value based on historical data are used for filling. Then, for the problem of inconsistent sampling frequencies of different data sources, precise time alignment operations are carried out. Among them, through advanced interpolation algorithms (such as spline interpolation) or down - sampling technical means, the load data, meteorological data, and distributed energy data are unified to the same time node, enabling effective fusion and analysis of multi - source data in the time dimension. Finally, the pre - processed data realizes the unified management and invocation of various data in the power grid resource business center, and various data are stored in a distributed database for subsequent efficient access and real - time update.

[0039] In a specific embodiment of the present invention, in order to meet the effectiveness and convenience requirements of different data operations, the model strategy, access mechanism and storage method of data are designed in each link of model development, sample construction, model training, data production, circulation, integration, analysis application, archiving and extinction. The data framework of the system is as follows: Figure 3 shown.

[0040] In an embodiment of the present invention, the edge-side collected data and business system data are preprocessed and labeled to form sample data that can be used for training and are collected in a sample library. Data from different business systems can be transferred by file or synchronized with directories, and the system uses sample data shared by all parties to uniformly organize model training. The model is mainly generated by the artificial intelligence platform (training environment) and imported into the model library. The development of a common prediction service model for power source and load at multiple spatiotemporal scales based on a pre-trained model is achieved by utilizing the artificial intelligence platform sample data, model library algorithm, training platform, and operation platform. The marketing, power distribution and other business fields call the model to push relevant business scenario data to the model, and the model returns the prediction result data.

[0041] An embodiment of the present invention proposes a power source and load prediction method based on multi-temporal and spatial data fusion. By integrating a temporal fusion transformer model and combining multi-source heterogeneous data such as power load, meteorology, and distributed energy, cross-temporal and cross-scale power load prediction is achieved. In a specific exemplary technical solution: First, power load, meteorological data, distributed energy generation data, and time features (such as seasons, holidays, timestamps, etc.) are mapped into high-dimensional feature vectors. The embedded feature vectors can retain the complex information of time-series data, providing a basis for subsequent time-series dependence analysis. Then, a gating module is introduced to filter out unnecessary components in the architecture, providing adaptive depth and network complexity to adapt to various datasets and scenarios, and solving the problem of multi-dimensional data in multi-temporal and spatial domains. Then, a variable selection network is used to further optimize the input data. This network is based on the attention mechanism and can automatically select the input variables most relevant to the current prediction task at each time step, reducing the interference of irrelevant information and improving the computational efficiency and prediction accuracy of the model. Multi-scale spatio-temporal dependence features are constructed through a static covariate encoder. This encoder uses the local perception and pooling operations of a convolutional neural network (CNN) to capture data features at different time scales and spatial dimensions and integrates them into context variables reflecting load fluctuations and meteorological changes. This vector integrates the dependence relationships of short-term and long-term time-series data, providing rich spatio-temporal information for load prediction. Finally, a sequence-to-sequence layer is used to process local information, and a temporal self-attention decoder is used to learn long-term patterns. The self-attention mechanism can automatically capture the correlations between different positions in the input sequence without pre-setting a fixed window size, thus better learning long-term time-series patterns. Through this unique structural design, the model can accurately identify the impact of meteorological changes on the load at different time scales, achieve cross-time-scale prediction fusion, and greatly improve the accuracy and reliability of the prediction.

[0042] In the technical solution disclosed in the embodiment of the present invention, the data flow is: raw data → variable selection network → static covariate encoder → time series processing module → temporal self-attention decoder → prediction result. Among them, the variable selection network provides concise key features for the static covariate encoder, reducing data dimension and noise interference. The static covariate encoder fuses the key features with multi-scale spatio-temporal dependence information to provide rich context for the time series processing module. The time series processing module captures local time series dependence and provides basic time series information for the temporal self-attention decoder. The temporal self-attention decoder uses the self-attention mechanism to combine local and global time series information to generate accurate predictions. The technical solution of the embodiment of the present invention can be dynamically adjusted according to the spatio-temporal characteristics of the input data. For example, during holidays or abnormal meteorological conditions, the variable selection network automatically increases the weights of relevant features; the static covariate encoder adjusts the convolution kernel size of the CNN to adapt to data of different spatial scales; the temporal self-attention decoder adjusts the number of attention heads to focus on patterns of different time scales.

[0043] In the preferred technical solution of the embodiment of the present invention, in order to ensure the accuracy of source-load prediction and the self-adaptability of the model, the system introduces a dynamic feedback mechanism to continuously optimize the model. When the model is migrated and deployed, based on the pre-trained model trained with existing large-scale historical power load data, meteorological data, and distributed energy data, by analyzing the similarity between the power load data in the new scenario and the existing scenarios, the most relevant input features are selected, such as historical load data, meteorological data, etc. In the case of high similarity, the intermediate layer features of the source model can be directly reused without retraining all layers. For the small amount of labeled data in the new scenario, the pre-trained model is fine-tuned. By adapting to the spatio-temporal features in different scenarios, the system will automatically adjust the model parameters and retrain the model according to the actual source-load situation and the error of the prediction result to optimize the prediction performance. Especially when significant changes occur in meteorological conditions and load demands, the system can quickly adjust the prediction result through real-time data update to ensure the real-time and flexibility of power dispatch.

[0044] Data at different time scales has different spatio-temporal characteristics. Through migration and deployment, the pre-trained model can be switched among intraday, short-term, and ultra-long-term prediction tasks. For example, intraday load prediction mainly relies on real-time data such as meteorological changes and instantaneous load fluctuations, while ultra-long-term load prediction is more inclined to be modeled through trend analysis. Through transfer learning, the model can be quickly switched between different time scales to improve the overall efficiency of the prediction system. The prediction results include intraday fluctuations, short-term load demands, and ultra-short-term load trend predictions. These prediction results will be output to the dispatching center in real time through the control cloud platform and the distributed resource aggregation platform for power dispatching personnel to formulate grid resource allocation plans and assist in power market transactions and energy dispatching.

[0045] In the embodiments of the present invention, based on the prediction service process, the overall system architecture is designed, and the business objectives, user requirements, and system functions of the service are clearly defined to ensure the understandability and usability of the system. The overall system architecture is shown in Figure 4 and can be divided into a basic layer, a framework layer, a model layer, and an application layer from bottom to top. Each layer complements each other to jointly build a complete, efficient, and stable prediction system.

[0046] Specifically by way of example, the basic layer provides the necessary computing resources, data resources, and storage resources for the entire prediction system. It uses a high-performance computing cluster for model training and inference, ensuring computational efficiency. In terms of data resources, it not only covers historical power load data from the power grid operation monitoring system, which details the power consumption at different time nodes, but also includes rich meteorological data from the enterprise-level meteorological data service center, such as parameters like temperature, humidity, wind speed, sunlight, rainfall, etc. In addition, distributed energy data is also incorporated, providing a key basis for analyzing the impact of distributed energy on power generation and load. The efficient storage system adopts a distributed storage architecture, such as object storage based on Ceph and distributed file systems like Hive, which can achieve fast data reading and writing and highly reliable storage, ensuring the integrity and availability of data throughout the system life cycle.

[0047] Specifically by way of example, the framework layer provides a stable and efficient operating environment for the entire prediction system. It adopts scheduling and control services Kubernetes (K8s) and Docker container technology to ensure the high availability and easy scalability of the system. In addition, the system also integrates various artificial intelligence-related algorithm computing frameworks, such as TensorFlow, PyTorch, etc., facilitating model training and inference.

[0048] Specific example explanatorily, the model layer is the core part of the prediction system of the present invention and is composed of a pre-trained model (i.e., the pre-trained model for the common prediction service of multi-temporal and multi-spatial scales of power source and load) and a series of dedicated fine-tuning models based on technologies such as LoRA (Low-Rank Adaptation) and SFT (Soft Prompt Tuning). Among them, the pre-trained model is trained on a large amount of historical data to extract the common characteristics of clean energy, providing a basis for the subsequent fine-tuning models; the fine-tuning models are optimized for specific application scenarios and prediction tasks. In a certain local power system, considering the local unique energy structure and electricity consumption habits, by adjusting the parameters and structure of the model, the model can better adapt to the characteristics of local power source and load changes, thus significantly improving the prediction accuracy. In addition, during the model training process, transfer learning technology is also adopted to transfer the effective knowledge and experience learned in other similar scenarios to the current model, accelerating the convergence speed of the model and improving its performance.

[0049] Specific example explanatorily, the application layer integrates core functional modules such as wind power prediction, photovoltaic power prediction, carrying capacity assessment, and overload prediction, and also incorporates an intelligent dispatching decision support module. Based on real-time prediction data and the grid operation status, this module automatically analyzes and generates the optimal power dispatching plan, including the priority ranking of clean energy generation, the charge and discharge strategies of energy storage systems, and demand-side response measures, etc., to achieve precise matching of energy supply and demand and stable and efficient operation of the power grid. In addition, a user interaction interface is developed to enable non-technical users to intuitively understand the prediction results and management strategies of clean energy, enhancing the usability and popularity of the system.

[0050] Regarding the hardware support and device implementation, the embodiment of the present invention also provides a hardware support device for the multi-temporal and multi-spatial scale prediction system of power source and load. This device provides multi-dimensional support services, such as Figure 5 shown.

[0051] In summary, the cross - time - space and cross - scale prediction system disclosed in the embodiments of the present invention is not only applicable to the load prediction of traditional power grids, but also capable of coping with complex scenarios after the access of distributed energy. For example, the system can be used in areas with a high proportion of new energy to accurately predict the power generation of distributed energy and its impact on the overall load, helping the power grid dispatching center to make optimized decisions. In addition, the system can also be applied to the power market trading platform to optimize market trading strategies and improve the utilization rate of power resources by predicting electricity prices, load demands, and new energy output. In the technical solutions of the embodiments of the present invention: First, historical power load, meteorological data, etc. are selected to construct a basic pre - training model, which is preliminarily trained under different power system scenarios to form a model library. Then, according to real - time requirements, the corresponding pre - training model is selected for transfer learning to quickly adapt to new prediction scenarios. By adjusting the model parameters, it can better reflect the characteristics of specific time scales and space scales, thereby achieving accurate load prediction. This method significantly improves the applicability and accuracy of the prediction model and effectively supports the dynamic dispatching and optimization of power systems.

[0052] The following is the device embodiment of the present invention, which can be used to execute the method embodiment of the present invention. For details not disclosed in the device embodiment, please refer to the method embodiment of the present invention.

[0053] Please refer to Figure 6 , in the embodiments of the present invention, a power source - load multi - time - space - scale commonality prediction system is provided, including: A model selection module, configured to select a pre - training model from a pre - constructed model library based on a selected application scenario and a prediction task; A model fine - tuning module, configured to use the selected pre - training model as a basic model and use the labeled data corresponding to the selected application scenario and prediction task as learning samples for transfer learning to obtain a fine - tuned power source - load prediction model. When performing transfer learning, the model parameters of the pre - training model are adjusted based on the source - load situation and prediction result error of the labeled data; A model prediction module, configured to perform prediction based on the selected application scenario and the prediction task, and use the obtained power source - load prediction model to obtain a prediction result; Among them, the pre - training models stored in the pre - constructed model library are all power source - load pre - training models trained based on existing large - scale historical power load data, meteorological data, and distributed energy data.

[0054] In one embodiment of the present invention, a computer device is provided. The computer device includes a processor and a memory. The memory is used to store a computer program, and the computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function. The processor described in the embodiments of the present invention can be used to execute the operations of the power source-load multi-temporal and spatial scale commonality prediction method.

[0055] In one embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device and is used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and, of course, the extended storage medium supported by the computer device. The computer-readable storage medium provides a storage space, and the operating system of the terminal is stored in this storage space. And, one or more instructions suitable for being loaded and executed by the processor are also stored in this storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed random access memory (RAM), or a non-volatile memory, such as at least one disk memory. One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the power source-load multi-temporal and spatial scale commonality prediction method in the above embodiments.

[0056] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) that contain computer-usable program code.

[0057] The present application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0058] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0059] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are performed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific implementation manners of the present invention. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention.

Claims

1. A common prediction method for power sources and loads at multiple spatio-temporal scales, characterized in that, It includes the following steps: Based on the selected application scenario and prediction task, select a pre-trained model from a pre-constructed model library; Use the selected pre-trained model as the base model, and use the labeled data corresponding to the selected application scenario and prediction task as the learning sample for transfer learning to obtain a fine-tuned power source and load prediction model; among them, when performing transfer learning, adjust the model parameters of the pre-trained model based on the source and load situation and prediction result error of the labeled data; Based on the selected application scenario and prediction task, use the obtained power source and load prediction model to make a prediction and obtain a prediction result; Among them, the pre-trained models stored in the pre-constructed model library are all power source and load pre-trained models obtained by training based on existing large-scale historical power load data, meteorological data, and distributed energy data.

2. The multi-temporal and multi-spatial scale common prediction method for power source and load according to claim 1, characterized in that The pre-trained model includes a data processing module, a gating module, and a multi-temporal and multi-spatial scale prediction module; among them, The data processing module is used to map the input data into an original high-dimensional feature vector; The gating module is used to filter redundant features in the original high-dimensional feature vector to obtain a processed high-dimensional feature vector; The multi-temporal and multi-spatial scale prediction module includes a variable selection network, a static covariate encoder, a time series processing module, and a time self-attention decoder; among them, the variable selection network is used to input the processed high-dimensional feature vector, and based on the attention mechanism, select the input features most relevant to the current prediction task at each time step, and output the screened key features and their weights; the static covariate encoder is used to input the screened key features and their weights, and use the local perception and pooling operations of the convolutional neural network to construct multi-scale spatio-temporal dependence features, and output a context vector that fuses multi-scale spatio-temporal dependence; the time series processing module is used to input the context vector that fuses multi-scale spatio-temporal dependence, perform local information processing and output time series data that retains local time series dependence; the time self-attention decoder is used to input the time series data that retains local time series dependence, capture the correlation between different positions in the sequence through the time self-attention mechanism, and combine local and global time series information to output a prediction result.

3. The multi-temporal and multi-spatial scale common prediction method for power source and load according to claim 2, characterized in that In the data processing module, the step of mapping the input data into an original high-dimensional feature vector includes: Obtain the original data; among them, the original data includes power load data, meteorological data, distributed energy generation data, and preset time features; Preprocess the original data to obtain preprocessed data; among them, the steps of preprocessing include: filling in the missing parts of the power load data and meteorological data, and unifying the power load data, meteorological data, and distributed energy generation data to the same time node; Perform mapping processing based on the preprocessed data to obtain an embedded original high-dimensional feature vector.

4. The multi-temporal and multi-spatial scale common prediction method for power source and load according to claim 3, characterized in that The power load data includes historical power load data of multiple time scales divided according to the collection frequency; The meteorological data includes one or more of temperature, humidity, wind speed, illumination, and rainfall, and is collected at minute-level, hour-level, and daily-level time scales; The preset time features include one or more of season, holiday, and timestamp.

5. A common prediction method for power source and load with multiple spatio-temporal scales according to claim 2, characterized in that, The time self-attention decoder adopts multi-head self-attention, and each head focuses on different preset time patterns.

6. A common prediction method for power source and load in multiple spatio-temporal scales according to claim 1, characterized in that, In the step of obtaining the prediction result by using the obtained power source and load prediction model based on the selected application scenario and prediction task, the obtained prediction result is applicable to wind power prediction, photovoltaic prediction, bearing capacity assessment, heavy overload prediction, or intelligent dispatching decision-making.

7. A power source and load multi - spatio - temporal scale commonality prediction system, characterized in that, It includes: A model selection module, configured to select a pre-trained model from a pre-constructed model library based on the selected application scenario and prediction task; A model fine-tuning module, configured to use the selected pre-trained model as the base model, and use the labeled data corresponding to the selected application scenario and prediction task as learning samples for transfer learning to obtain a fine-tuned power source and load prediction model; wherein, when performing transfer learning, the model parameters of the pre-trained model are adjusted based on the source and load conditions and prediction result errors of the labeled data; A model prediction module, configured to perform prediction by using the obtained power source and load prediction model based on the selected application scenario and prediction task to obtain a prediction result; Among them, the pre-trained models stored in the pre-constructed model library are all power source and load pre-trained models obtained by training based on existing large-scale historical power load data, meteorological data, and distributed energy data.

8. A common prediction system for power source and load with multiple spatio-temporal scales according to claim 7, characterized in that, The pre-trained model includes a data processing module, a gating module, and a multi-spatio-temporal scale prediction module; wherein, The data processing module is configured to map the input data into an original high-dimensional feature vector; The gating module is configured to filter redundant features in the original high-dimensional feature vector to obtain a processed high-dimensional feature vector; The multi - spatio - temporal scale prediction module includes a variable selection network, a static covariate encoder, a time series processing module, and a temporal self - attention decoder. Among them, the variable selection network is used to input the processed high - dimensional feature vector, select the input features most relevant to the current prediction task at each time step based on the attention mechanism, and output the filtered key features and their weights. The static covariate encoder is used to input the filtered key features and their weights, and use the local perception and pooling operations of the convolutional neural network to construct multi - scale spatio - temporal dependence features, and output the context vector integrating multi - scale spatio - temporal dependence. The time series processing module is used to input the context vector integrating multi - scale spatio - temporal dependence, perform local information processing, and output the time series data retaining local time series dependence. The temporal self - attention decoder is used to input the time series data retaining local time series dependence, capture the correlation between different positions in the sequence through the temporal self - attention mechanism, and output the prediction result by combining local and global time series information.

9. The multi - spatio - temporal scale commonality prediction system for power source and load according to claim 8, wherein In the data processing module, the steps of mapping the input data into the original high - dimensional feature vector include: Obtain the original data. Among them, the original data includes power load data, meteorological data, distributed energy generation data, and preset time features. Pre - process the original data to obtain the pre - processed data. Among them, the steps of pre - processing include: filling in the missing parts of the power load data and meteorological data, and unifying the power load data, meteorological data, and distributed energy generation data to the same time node. Perform mapping processing based on the pre - processed data to obtain the embedded original high - dimensional feature vector.

10. The multi - spatio - temporal scale commonality prediction system for power source and load according to claim 9, wherein The power load data includes historical power load data of multiple time scales divided according to the acquisition frequency. The meteorological data includes one or more of temperature, humidity, wind speed, light, and rainfall, and is collected at minute - level, hour - level, and daily - level time scales. The preset time features include one or more of season, holiday, and timestamp.

11. The multi - spatio - temporal scale commonality prediction system for power source and load according to claim 8, wherein The temporal self - attention decoder uses multi - head self - attention, and each head focuses on different preset time patterns.

12. The multi - spatio - temporal scale commonality prediction system for power source and load according to claim 7, characterized in that, In the step of obtaining the prediction result by using the obtained power source and load prediction model based on the selected application scenario and prediction task, the obtained prediction result is applicable to wind power prediction, photovoltaic prediction, carrying capacity assessment, heavy overload prediction, or intelligent dispatching decision.

13. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multi - spatio - temporal scale commonality prediction method for power source and load as described in any one of claims 1 to 6.

14. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi - spatio - temporal scale commonality prediction method for power source and load as described in any one of claims 1 to 6.

Images