Park load power prediction method and device and readable storage medium

Abstract

The invention provides a park load power prediction method and device and a readable storage medium. The park load power prediction method comprises the steps of obtaining historical multi-source data of a park; obtaining time sequence data based on the historical multi-source data; sampling the time sequence data based on multiple sampling precisions, and determining a sampling data set of multiple time scales; based on the sampling data set and a preset model, obtaining a plurality of load power prediction sub-models corresponding to the multiple time scales; obtaining a target time period, and obtaining a load power prediction sequence of each load power prediction sub-model in the target time period according to the historical data length, the single prediction period number and the time sequence data corresponding to each load power prediction sub-model; performing probability fusion according to the plurality of load power prediction sequences to obtain a load power prediction integrated model; and outputting a load power fusion prediction value corresponding to the target time period according to the load power prediction integration model. Multiple sampling accuracies can increase the accuracy of model prediction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method, apparatus, and readable storage medium for predicting the load power of a park. Background Technology

[0002] Among related technologies, load power forecasting is a core means of optimizing the operation of integrated energy parks and virtual power plants, aiming to improve dispatch efficiency and economy by accurately predicting electricity demand. However, deep learning methods typically rely on data analysis at a single time scale, making it impossible for the model to collaboratively capture both the macro-long-term trend and micro-short-term fluctuations of load, resulting in insufficient prediction accuracy and difficulty in meeting the needs of refined dispatching. Summary of the Invention

[0003] The present invention aims to solve at least one of the technical problems existing in the prior art or related art.

[0004] Therefore, the first aspect of the present invention proposes a method for predicting the load power of a park.

[0005] A second aspect of the present invention provides a campus load power prediction device.

[0006] A third aspect of the present invention provides another campus load power prediction device.

[0007] The fourth aspect of this application proposes a readable storage medium.

[0008] In view of this, a first aspect of the present invention provides a method for predicting the load power of a park, comprising: acquiring historical multi-source data of the park, wherein the historical multi-source data includes at least load power-related data; acquiring time-series data based on the historical multi-source data; sampling the time-series data based on multiple sampling accuracies to determine a multi-time-scale sampling dataset, wherein the multiple sampling accuracies are associated with multiple time scales; acquiring multiple load power prediction sub-models corresponding to the multiple time scales based on the sampling dataset and a preset model, wherein each load power prediction sub-model is a hybrid model combining a convolutional neural network and a gated recurrent unit; acquiring a target time period, and acquiring a load power prediction sequence for each load power prediction sub-model within the target time period based on the historical data length, single prediction period number, and time-series data corresponding to each load power prediction sub-model; performing probabilistic fusion of multiple load power prediction sequences to obtain a load power prediction ensemble model; and outputting a load power fusion prediction value corresponding to the target time period based on the load power prediction ensemble model.

[0009] This application acquires historical multi-source data from the park and constructs a multi-timescale sampling dataset, enabling the model to simultaneously capture both micro-fluctuations and macro-trends in load. This avoids the problem of insufficient prediction accuracy at a single scale. Understandably, this application expands the range of data sources and increases data dimensionality to reduce prediction errors. It is important to emphasize that this application samples time-series data according to different sampling accuracies. Since different sampling accuracies correspond to different sampling frequencies, such as low-frequency, medium-frequency, and high-frequency, sampling datasets at different time scales are obtained based on multiple sampling accuracies. At the data level, through a hierarchical sampling method, the sampled datasets accurately reflect the changes in coincident power under different conditions.

[0010] Meanwhile, based on the hybrid model combining convolutional neural networks (CNN) and gated recurrent units (GRU), it is possible to learn the load change characteristics at different scales. By leveraging the local feature extraction capability of CNN and the temporal modeling advantage of GRU, the feature representation capability is improved.

[0011] Then, a rolling prediction mechanism is used to iteratively generate prediction sequences based on the historical data length and single prediction period of each sub-model, which effectively controls error accumulation. Finally, probabilistic fusion is used to integrate multi-scale prediction results, and weights are assigned according to the posterior probability of each model on the validation set to achieve uncertainty quantification. Ultimately, the load prediction accuracy within the target period is improved, and the prediction accuracy is improved compared to a single model. At the same time, the prediction robustness and engineering practicality under complex scenarios such as extreme weather are enhanced.

[0012] Specifically, this application, based on a pre-defined CNN-GRU model architecture, trains sub-models at different sampling accuracies (e.g., 5 minutes, 15 minutes, and 1 hour), enabling each sub-model to exhibit a tiered predictive capability difference for high-frequency, mid-frequency, and low-frequency load fluctuations: the high-frequency model accurately captures short-term fluctuations, the mid-frequency model balances transitional trends, and the low-frequency model stably reflects long-term patterns. This difference stems from each model's focused learning of time-scale-specific features. Furthermore, the ensemble model, which integrates these models with significant capability differences but good performance in their respective domains, exhibits slightly higher predictive accuracy and stronger robustness. This is achieved by using Bayesian model averaging (BMA) for probability fusion, dynamically weighting the posterior probabilities of each sub-model, thereby combining the detail sensitivity of high-frequency models with the trend stability of low-frequency models during integration, reducing the uncertainty of individual models.

[0013] In some technical solutions of this application, obtaining time-series data based on historical multi-source data includes: performing data cleaning, time-series alignment and feature engineering on historical multi-source data to obtain time-series data.

[0014] In this technical solution, the present application obtains time-series data by performing data cleaning, time-series alignment, and feature engineering on historical multi-source data, thereby realizing a systematic data preprocessing process that can improve data quality. It enhances the integrity of the original data, specifically increasing it from approximately 90% to 99.7%, providing a reliable data foundation for subsequent model training.

[0015] This application addresses the direct impact of data quality on the performance of machine learning models. It considers that missing and outlier values ​​can lead to incorrect model learning, while temporal misalignment can disrupt causal relationships between features. Therefore, this application employs PCHIP interpolation to preserve data distribution characteristics, combines forward and backward padding to handle boundary issues, and finally establishes accurate temporal relationships through a unified time index. This series of operations ensures the consistency and integrity of the time-series data.

[0016] In some technical solutions of this application, data cleaning, time-series alignment and feature engineering are performed on historical multi-source data to obtain time-series data. This includes: interpolating missing values ​​in historical multi-source data using interpolation methods to obtain preprocessed data, wherein missing values ​​at the beginning and end of the data sequence are interpolated using forward padding or backward padding methods; and unifying preprocessed data with different sampling precisions onto the same time index to obtain time-series data.

[0017] In this technical solution, the present application uses interpolation methods to impute missing values ​​in historical multi-source data to obtain preprocessed data. Specifically, for missing values ​​at the beginning and end of the data sequence, forward imputation or backward imputation methods are used. Preprocessed data with different sampling precisions are unified onto the same time index to obtain time-series data. That is, the processing effect of different types of missing values ​​is optimized through differentiated interpolation strategies. This avoids the numerical overflow problem caused by a single interpolation method at data boundaries. Furthermore, based on the interpolation processing method of this application, the interpolation error can be reduced by approximately 40%.

[0018] In some technical solutions of this application, multiple sampling accuracies include at least one low-frequency sampling accuracies, at least one medium-frequency sampling accuracies, and at least one high-frequency sampling accuracies.

[0019] In this technical solution, multiple sampling accuracies include at least one low-frequency sampling accuracies, at least one mid-frequency sampling accuracies, and at least one high-frequency sampling accuracies. This three-scale sampling architecture achieves a panoramic capture of load characteristics. Specifically, it can improve the model's interpretability of load changes by 35%, while reducing prediction errors by approximately 25% in abrupt changes such as extreme weather.

[0020] This application is based on the multi-scale characteristics of load fluctuations. Specifically, high-frequency changes reflect instantaneous behaviors such as equipment start-up and shutdown, medium-frequency changes reflect daily operating cycles, and low-frequency changes correspond to seasonal trends. A single scale can only capture some features, while the three-level sampling of 5 minutes, 15 minutes, and 1 hour set up in this application forms a complete observation system from micro to macro, enabling the model to simultaneously grasp short-term fluctuation patterns and long-term trend changes.

[0021] In some technical solutions of this application, obtaining the load power prediction sequence of each load power prediction sub-model within the target time period includes: dividing the target time period according to the single prediction period corresponding to each load power prediction sub-model, and determining the number of rolling predictions required for prediction; extracting an initial historical data window from the time series data based on the length of historical data; performing the first forward prediction using the historical data window to obtain the first prediction sequence; using the result of the first prediction sequence as the updated historical data, updating the historical data window, and deleting the historical data corresponding to the historical time furthest from the current time to keep the window length of the historical data window unchanged; repeating the prediction and window update steps until the rolling prediction of the rolling prediction number is completed, and splicing each prediction sequence in sequence to determine the load power prediction sequence.

[0022] In this technical solution, the target time period is divided according to the number of single prediction periods corresponding to each load power prediction sub-model, determining the number of rolling predictions required. Based on the length of historical data, an initial historical data window is extracted from the time series data. The historical data window is used for the first forward prediction to obtain the first prediction sequence. The result of the first prediction sequence is used as the updated historical data to update the historical data window, and the historical data corresponding to the historical time furthest from the current time is deleted to keep the window length unchanged. The prediction and window update steps are repeated until the rolling prediction is completed, and each prediction sequence is concatenated in sequence to determine the load power prediction sequence. This decomposes the long-term prediction task into multiple short-term prediction stages, reducing accumulated errors and minimizing the propagation of uncertainty in model predictions. The data window is updated immediately after each prediction with the latest results, forming a closed-loop control system of "prediction-verification-correction," effectively preventing the exponential amplification of errors. Rolling prediction enables the model to continuously absorb the latest operational data, improving the response speed to load fluctuations and significantly enhancing the model's online learning capability.

[0023] In some technical solutions of this application, in each rolling forecast, the historical data window used for forecasting includes preprocessed real historical data, or forecast data generated from the previous rolling forecast, or a combination of real historical data and forecast data generated from the previous rolling forecast.

[0024] In this technical solution, in each rolling forecast, the historical data window used for forecasting contains preprocessed real historical data, forecast data generated from the previous rolling forecast, or a combination of real historical data and forecast data generated from the previous rolling forecast. In other words, the accuracy and real-time performance of the forecast are balanced through a flexible data window combination strategy.

[0025] In some of the technical solutions of this application, the sampling time interval corresponding to low frequency sampling accuracy is 1 hour, the sampling time interval corresponding to medium frequency sampling accuracy is 15 minutes, and the sampling time interval corresponding to high frequency sampling accuracy is 5 minutes.

[0026] In this technical solution, the sampling time interval corresponding to low-frequency sampling accuracy is 1 hour, the sampling time interval corresponding to medium-frequency sampling accuracy is 15 minutes, and the sampling time interval corresponding to high-frequency sampling accuracy is 5 minutes. Optimal balance between computational efficiency and prediction accuracy is achieved through precisely set sampling intervals.

[0027] In some of the technical solutions of this application, the historical multi-source data also includes the grid connection power data of the park's grid connection points, the energy storage charging and discharging power data, the power generation power data, the park's historical meteorological data, meteorological forecast data, operating condition data, and the park's production scheduling information.

[0028] In this technical solution, historical multi-source data also includes grid connection power data of the park's grid connection points, energy storage charging and discharging power data, power generation power data, historical meteorological data of the park, meteorological forecast data, operating condition data, and production scheduling information of the park. Increasing the types and scope of data can improve the accuracy and effectiveness of the initial data, thereby further increasing the prediction accuracy during subsequent model learning. At the same time, the intervention of multi-dimensional data can increase the prediction accuracy of the model in special scenarios.

[0029] For example, meteorological data is directly related to air conditioning load (every 1°C change in temperature affects load by 2%-5%), production scheduling information reflects planned capacity changes, and grid connection power reflects grid interaction needs. By placing load power in a causal network composed of multi-source data, the model can distinguish between random fluctuations and regular changes, improving its generalization ability to complex scenarios.

[0030] This application provides a park load power prediction device, including a first acquisition module, a second acquisition module, a first determination module, a third acquisition module, a fourth acquisition module, a fifth acquisition module, and a first output module. The first acquisition module is used to acquire historical multi-source data of the park, wherein the historical multi-source data includes at least load power related data. The second acquisition module is used to acquire time-series data based on the historical multi-source data. The first determination module is used to sample the time-series data based on multiple sampling accuracies to determine a multi-time-scale sampling dataset, wherein the multiple sampling accuracies are associated with multiple time scales. The third acquisition module is used to acquire data related to the sampling dataset and a preset model. The system comprises multiple load power prediction sub-models corresponding to multiple time scales, each of which is a hybrid model combining convolutional neural networks and gated recurrent units. The fourth acquisition module is used to acquire the target time period and, based on the historical data length, single prediction period number, and time series data corresponding to each load power prediction sub-model, acquires the load power prediction sequence of each load power prediction sub-model within the target time period. The fifth acquisition module is used to perform probabilistic fusion of multiple load power prediction sequences to acquire an integrated load power prediction model. The first output module is used to output the fused load power prediction value corresponding to the target time period based on the integrated load power prediction model.

[0031] This application provides a park load power prediction device, including a first acquisition module, a second acquisition module, a first determination module, a third acquisition module, a fourth acquisition module, a fifth acquisition module, and a first output module. The first acquisition module is used to acquire historical multi-source data of the park, wherein the historical multi-source data includes at least load power related data; the second acquisition module is used to acquire time-series data based on the historical multi-source data; the first determination module is used to sample the time-series data based on multiple sampling accuracies to determine a multi-time-scale sampling dataset, wherein the multiple sampling accuracies are associated with multiple time scales; the third acquisition module is used to acquire data related to multiple time scales based on the sampling dataset and a preset model. The system comprises multiple load power prediction sub-models corresponding to the scale, each of which is a hybrid model combining a convolutional neural network and a gated recurrent unit. The fourth acquisition module is used to acquire the target time period and, based on the historical data length, single prediction period, and time series data corresponding to each load power prediction sub-model, acquires the load power prediction sequence of each load power prediction sub-model within the target time period. The fifth acquisition module is used to perform probabilistic fusion based on multiple load power prediction sequences to acquire an integrated load power prediction model. The first output module is used to output the fused load power prediction value corresponding to the target time period based on the integrated load power prediction model.

[0032] This application provides a campus load power prediction device, including a processor and a memory. The memory stores programs or instructions, and when the processor executes the programs or instructions in the memory, it implements the steps of the campus load power prediction method as described in any of the above-described technical solutions. Therefore, the campus load power prediction device possesses all the beneficial effects of the campus load power prediction method described in any of the above-described technical solutions.

[0033] This application provides a readable storage medium on which a program or instructions are stored. When executed by a processor, the program or instructions implement the steps of the campus load power prediction method as described in any of the above-described technical solutions. Therefore, the readable storage medium possesses all the beneficial effects of the campus load power prediction method as described in any of the above-described technical solutions.

[0034] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0035] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0036] Figure 1 This is one of the flowcharts illustrating a method for predicting park load power according to an embodiment of the present invention;

[0037] Figure 2 This is a second schematic flowchart of a method for predicting the load power of a park according to an embodiment of the present invention;

[0038] Figure 3 This is a model network structure diagram according to an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of a multi-step rolling prediction logic according to an embodiment of the present invention;

[0040] Figure 5 This is one of the comparison diagrams of the load power prediction effect of the park load power prediction method according to an embodiment of the present invention;

[0041] Figure 6 This is the second comparison diagram of the load power prediction effect of the park load power prediction method according to an embodiment of the present invention;

[0042] Figure 7 This is a comparison chart of the prediction performance of an integrated model and a single model according to an embodiment of the present invention;

[0043] Figure 8This is one of the schematic block diagrams of a park load power prediction device according to an embodiment of the present invention;

[0044] Figure 9 This is a second schematic block diagram of a park load power prediction device according to an embodiment of the present invention. Detailed Implementation

[0045] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0047] The following reference Figures 1 to 9 This invention describes a method, apparatus, and readable storage medium for predicting campus load power according to some embodiments of the present invention.

[0048] like Figure 1 As shown in the embodiments of this application, a method for predicting the load power of a park is provided, the steps of which include:

[0049] Step 102: Obtain historical multi-source data of the park, wherein the historical multi-source data includes at least load power related data;

[0050] Step 104: Obtain time-series data based on historical multi-source data;

[0051] Step 106: Sample the time series data based on multiple sampling precipitates to determine a multi-time-scale sampling dataset, wherein the multiple sampling precipitates are associated with multiple time scales;

[0052] Step 108: Based on the sampled dataset and the preset model, obtain multiple load power prediction sub-models corresponding to multiple time scales. Each load power prediction sub-model is a hybrid model combining a convolutional neural network and a gated recurrent unit.

[0053] Step 110: Obtain the target time period. Based on the historical data length, single prediction period, and time series data corresponding to each load power prediction sub-model, obtain the load power prediction sequence of each load power prediction sub-model within the target time period.

[0054] Step 112: Perform probabilistic fusion based on multiple load power prediction sequences to obtain an integrated load power prediction model;

[0055] Step 114: Based on the load power prediction integration model, output the load power fusion prediction value corresponding to the target time period.

[0056] This application acquires historical multi-source data from the park and constructs a multi-timescale sampling dataset, enabling the model to simultaneously capture both micro-fluctuations and macro-trends in load. This avoids the problem of insufficient prediction accuracy at a single scale. Understandably, this application expands the range of data sources and increases data dimensionality to reduce prediction errors. It is important to emphasize that this application samples time-series data according to different sampling accuracies. Since different sampling accuracies correspond to different sampling frequencies, such as low-frequency, medium-frequency, and high-frequency, sampling datasets at different time scales are obtained based on multiple sampling accuracies. At the data level, through a hierarchical sampling method, the sampled datasets accurately reflect the changes in coincident power under different conditions.

[0057] Meanwhile, based on the hybrid model combining convolutional neural networks (CNN) and gated recurrent units (GRU), it is possible to learn the load change characteristics at different scales. By leveraging the local feature extraction capability of CNN and the temporal modeling advantage of GRU, the feature representation capability is improved.

[0058] Then, a rolling prediction mechanism is used to iteratively generate prediction sequences based on the historical data length and single prediction period of each sub-model, which effectively controls error accumulation. Finally, probabilistic fusion is used to integrate multi-scale prediction results, and weights are assigned according to the posterior probability of each model on the validation set to achieve uncertainty quantification. Ultimately, the load prediction accuracy within the target period is improved, and the prediction accuracy is improved compared to a single model. At the same time, the prediction robustness and engineering practicality under complex scenarios such as extreme weather are enhanced.

[0059] In one embodiment, by integrating multi-source data, the feature dimensions of load forecasting are expanded from single power data to more than 10 related features such as meteorological and operating conditions. This allows the model to capture the impact of key features such as temperature, humidity, and production plans on the load, effectively reducing the prediction error in special scenarios (such as extreme weather) by more than 30%. A piecewise cubic Hermite interpolating polynomial (PCHIP) interpolation combined with a forward / backward padding data cleaning strategy solves the problem of missing power data caused by the intermittency of photovoltaic power generation, improving data integrity to 99.7% and providing high-quality input for model training.

[0060] In one embodiment, the CNN-GRU hybrid architecture can reduce memory error while maintaining training speed, achieving short-term prediction accuracies of 94.36% (5 minutes), 93.88% (15 minutes), and 92.55% (1 hour) on the validation set. Here, 5 minutes, 15 minutes, and 1 hour refer to the specific sampling precision.

[0061] The CNN part extracts local correlation features of the load sequence through one-dimensional convolutional kernels, while the GRU part captures long-term dependencies in the load sequence, such as daily and weekly time patterns, through update and reset gate mechanisms, forming a complementary spatiotemporal feature set.

[0062] Specifically, this application, based on a pre-defined CNN-GRU model architecture, trains sub-models at different sampling accuracies (e.g., 5 minutes, 15 minutes, and 1 hour), enabling each sub-model to exhibit a tiered predictive capability difference for high-frequency, mid-frequency, and low-frequency load fluctuations: the high-frequency model accurately captures short-term fluctuations, the mid-frequency model balances transitional trends, and the low-frequency model stably reflects long-term patterns. This difference stems from each model's focused learning of time-scale-specific features. Furthermore, the ensemble model, which integrates these models with significant capability differences but good performance in their respective domains, exhibits slightly higher predictive accuracy and stronger robustness. This is achieved by using Bayesian model averaging (BMA) for probability fusion, dynamically weighting the posterior probabilities of each sub-model, thereby combining the detail sensitivity of high-frequency models with the trend stability of low-frequency models during integration, reducing the uncertainty of individual models.

[0063] In one embodiment, the solution of this application obtains the target time period. Based on the historical data length, single prediction period, and time series data corresponding to each load power prediction sub-model, the load power prediction sequence of each load power prediction sub-model within the target time period is obtained. That is, through a rolling prediction mechanism, the model can continuously use the latest prediction results to correct subsequent predictions, thereby reducing the cumulative error of the prediction task within the target time period. Specifically, the prediction sub-models corresponding to different time scales can adopt a differentiated rolling strategy. That is, the model with a sampling accuracy of 5 minutes rolls 12 times (predicting 2 hours each time), the model with a sampling accuracy of 15 minutes rolls 4 times (predicting 6 hours each time), and the model with a sampling accuracy of 1 hour predicts 24 hours at a time. By combining the prediction sequences corresponding to the models with different sampling accuracy, and then fusing them, the optimal balance between accuracy and efficiency can be achieved.

[0064] In one embodiment, during the process of obtaining an integrated load power prediction model by probabilistic fusion based on multiple load power prediction sequences, probabilistic fusion can be performed using the Bayesian model averaging method. The Bayesian model averaging method (BMA) can calculate posterior probability weights based on the historical prediction performance of each model, thereby achieving probabilistic fusion and providing uncertainty quantification indicators.

[0065] In this application, the rolling prediction method iteratively updates the input data through a sliding time window, enabling the model to continuously correct the prediction results based on the latest information, which is suitable for multi-step prediction tasks.

[0066] The multi-timescale approach in this application refers to simultaneously analyzing the variation patterns of load data at different time granularities (such as minutes, hours, and days), overcoming the limitations of single-scale analysis.

[0067] In some embodiments of this application, obtaining time-series data based on historical multi-source data includes: performing data cleaning, time-series alignment, and feature engineering on historical multi-source data to obtain time-series data.

[0068] In this embodiment, this application obtains time-series data by performing data cleaning, time-series alignment, and feature engineering on historical multi-source data, thereby realizing a systematic data preprocessing process that can improve data quality. The integrity of the original data is improved, specifically from approximately 90% to 99.7%, providing a reliable data foundation for subsequent model training.

[0069] This application addresses the direct impact of data quality on the performance of machine learning models. It considers that missing and outlier values ​​can lead to incorrect model learning, while temporal misalignment can disrupt causal relationships between features. Therefore, this application employs PCHIP interpolation to preserve data distribution characteristics, combines forward and backward padding to handle boundary issues, and finally establishes accurate temporal relationships through a unified time index. This series of operations ensures the consistency and integrity of the time-series data.

[0070] In some embodiments of this application, data cleaning, time-series alignment, and feature engineering are performed on historical multi-source data to obtain time-series data. This includes: interpolating missing values ​​in the historical multi-source data using interpolation methods to obtain preprocessed data; wherein, for missing values ​​at the beginning and end of the data sequence, forward padding or backward padding methods are used for interpolation; and preprocessed data with different sampling precisions are unified onto the same time index to obtain time-series data.

[0071] In this embodiment, this application uses interpolation methods to impute missing values ​​in historical multi-source data to obtain preprocessed data. Specifically, for missing values ​​at the beginning and end of the data sequence, forward padding or backward padding methods are used. Preprocessed data with different sampling accuracies are unified to the same time index to obtain time-series data. That is, the processing effect of different types of missing values ​​is optimized through differentiated interpolation strategies. This avoids the numerical overflow problem caused by a single interpolation method at data boundaries. Furthermore, based on the interpolation processing method of this application, the interpolation error can be reduced by approximately 40%.

[0072] Specifically, missing data in the middle section has reference points before and after it, making PCHIP interpolation suitable for maintaining curve smoothness; however, missing data at the beginning and end only has reference points on one side, and forcing the use of PCHIP would amplify the extrapolation error. Therefore, a more conservative strategy of forward / backward padding is adopted to ensure data continuity while avoiding overfitting.

[0073] In some embodiments of this application, the multiple sampling accuracies include at least one low-frequency sampling accuracies, at least one medium-frequency sampling accuracies, and at least one high-frequency sampling accuracies.

[0074] In this embodiment, the multiple sampling accuracies include at least one low-frequency sampling accuracies, at least one mid-frequency sampling accuracies, and at least one high-frequency sampling accuracies. This three-scale sampling architecture achieves a panoramic capture of load characteristics. Specifically, it can improve the model's interpretability of load changes by 35%, while reducing prediction errors by approximately 25% in abrupt changes such as extreme weather.

[0075] This application is based on the multi-scale characteristics of load fluctuations. Specifically, high-frequency changes reflect instantaneous behaviors such as equipment start-up and shutdown, medium-frequency changes reflect daily operating cycles, and low-frequency changes correspond to seasonal trends. A single scale can only capture some features, while the three-level sampling of 5 minutes, 15 minutes, and 1 hour set up in this application forms a complete observation system from micro to macro, enabling the model to simultaneously grasp short-term fluctuation patterns and long-term trend changes.

[0076] In some embodiments of this application, obtaining the load power prediction sequence of each load power prediction sub-model within the target time period includes: dividing the target time period according to the single prediction period corresponding to each load power prediction sub-model, and determining the number of rolling predictions required for prediction; extracting an initial historical data window from the time series data based on the historical data length; performing the first forward prediction using the historical data window to obtain the first prediction sequence; using the result of the first prediction sequence as the updated historical data, updating the historical data window, and deleting the historical data corresponding to the historical time furthest from the current time to keep the window length of the historical data window unchanged; repeating the prediction and window update steps until the rolling prediction of the number of rolling predictions is completed, and splicing each prediction sequence in sequence to determine the load power prediction sequence.

[0077] In this embodiment, the target time period is divided according to the number of single prediction periods corresponding to each load power prediction sub-model, and the number of rolling predictions required for prediction is determined. Based on the length of historical data, an initial historical data window is extracted from the time series data. The historical data window is used for the first forward prediction to obtain the first prediction sequence. The result of the first prediction sequence is used as the updated historical data to update the historical data window, and the historical data corresponding to the historical time furthest from the current time is deleted to keep the window length of the historical data window unchanged. The prediction and window update steps are repeated until the rolling prediction is completed for the required number of rolling predictions. Each prediction sequence is then concatenated in sequence to determine the load power prediction sequence. This decomposes the long-term prediction task into multiple short-term prediction stages, reducing accumulated errors and minimizing the propagation of uncertainty in model predictions. The data window is updated immediately after each prediction with the latest results, forming a closed-loop control system of "prediction-verification-correction," which effectively prevents the exponential amplification of errors. Rolling prediction enables the model to continuously absorb the latest operating data, improving the response speed to load changes and significantly enhancing the model's online learning capability.

[0078] For example, this application effectively controls the accumulation of errors in long-term forecasts through a dynamic window update mechanism. When the target period is 24 hours, the forecast task of the scheme is 24 hours. In actual forecasting, the longer the single forecast step, the greater the uncertainty. This application can divide the 24-hour target into multiple short periods (e.g., a 5-minute model makes 12 forecasts), forecasting only 2 hours each time and updating the input data based on the results. This ensures that the model always performs the next round of forecasting with the latest state, similar to a closed-loop control system that continuously corrects deviations, thereby significantly improving the stability of long-term forecasts and reducing the error to within 0.8%.

[0079] In some embodiments of this application, in each rolling prediction, the historical data window used for prediction contains preprocessed real historical data, or contains prediction data generated from the previous rolling prediction, or contains a combination of real historical data and prediction data generated from the previous rolling prediction.

[0080] In this embodiment, in each rolling prediction, the historical data window used for prediction contains preprocessed real historical data, or prediction data generated from the previous rolling prediction, or a combination of real historical data and prediction data generated from the previous rolling prediction. In other words, the accuracy and real-time performance of the prediction are balanced through a flexible data window combination strategy.

[0081] Specifically, in rolling forecasting, data combination and iteration are used to achieve rolling forecasting, which improves the model's response speed by about 30% when making continuous forecasts and enhances its adaptability to sudden events.

[0082] Historical data is accurate but outdated, while forecast data is timely but contains errors. By reasonably combining the two (e.g., using measured data for the first 4 hours and forecast data for the next 20 hours), the reliability of near-term forecasts is ensured while achieving sufficient coverage of the future. This gradual transition strategy conforms to the positive correlation between forecast uncertainty and time.

[0083] In some embodiments of this application, the sampling time interval corresponding to low-frequency sampling accuracy is 1 hour, the sampling time interval corresponding to medium-frequency sampling accuracy is 15 minutes, and the sampling time interval corresponding to high-frequency sampling accuracy is 5 minutes.

[0084] In this embodiment, the sampling time interval corresponding to low-frequency sampling accuracy is 1 hour, the sampling time interval corresponding to medium-frequency sampling accuracy is 15 minutes, and the sampling time interval corresponding to high-frequency sampling accuracy is 5 minutes. This precisely set sampling interval achieves an optimal balance between computational efficiency and prediction accuracy.

[0085] Specifically, actual tests show that this combination of time intervals improves the model training speed by about 40%, while the key feature capture completeness reaches over 95%.

[0086] For example, a 5-minute interval can capture the typical start-stop cycle (3-10 minutes) of equipment such as elevators and air conditioners, a 15-minute interval corresponds to the local power grid adjustment cycle, and a 1-hour interval matches the base load change cycle. This interval setting avoids data redundancy and noise amplification caused by too small an interval, while preventing feature loss caused by too large an interval, so that models at each scale can focus on the most representative change patterns.

[0087] In some embodiments of this application, the historical multi-source data also includes the grid connection power data of the park's grid connection points, the energy storage charging and discharging power data, the power generation power data, the park's historical meteorological data, meteorological forecast data, operating condition data, and the park's production scheduling information.

[0088] In this embodiment, the historical multi-source data also includes the grid connection power data of the park's grid connection points, energy storage charging and discharging power data, power generation power data, historical meteorological data of the park, meteorological forecast data, operating condition data, and production scheduling information of the park. Increasing the types and scope of data can improve the accuracy and effectiveness of the initial data, thereby further increasing the prediction accuracy during subsequent model learning. At the same time, the intervention of multi-dimensional data can increase the prediction accuracy of the model in special scenarios.

[0089] For example, meteorological data is directly related to air conditioning load (every 1°C change in temperature affects load by 2%-5%), production scheduling information reflects planned capacity changes, and grid connection power reflects grid interaction needs. By placing load power in a causal network composed of multi-source data, the model can distinguish between random fluctuations and regular changes, improving its generalization ability to complex scenarios.

[0090] like Figure 2 As shown, the park load power prediction method of this application also includes:

[0091] Step 202: Acquire multi-source data, including grid connection point power data, energy storage power data, power generation data, historical meteorological data, and production scheduling data;

[0092] Step 204, data preprocessing and feature engineering, including data imputation, feature dimensionality reduction, dimensionless transformation, data segmentation and feature enhancement;

[0093] Step 206, Dataset Construction, including 15-minute resampled dataset, 5-minute resampled dataset, and 1-hour resampled dataset;

[0094] Step 208: Based on different datasets, train composite models of different convolutional neural networks and gated recurrent units;

[0095] Step 210: Obtain the integrated prediction model;

[0096] Step 212, Model Invocation and Prediction.

[0097] Among them, multi-source data refers to load-related data obtained from various sources such as grid-connected power, energy storage power, power generation, historical weather, and production scheduling plans; resampling involves sampling time-series data at different time intervals, such as 5 minutes, 15 minutes, and 1 hour, to form a multi-time-scale dataset; the hybrid model (Convolutional Neural Network-Gated Recurrent Unit, CNN-GRU) is a hybrid model that integrates convolutional neural networks and gated recurrent units to collaboratively extract spatial features and temporal dependencies; step 210, obtaining the Bayesian Model Averaging (BMA) method used by the ensemble prediction model, is a method based on posterior probability weighting and fusion of multi-model prediction results. Figure 2The operating principle begins with the acquisition of multi-source data, followed by data preprocessing and feature engineering, including data completion, feature dimensionality reduction, dimensionless transformation, data segmentation, and feature enhancement. The preprocessed data is then resampled into datasets at different scales, such as 5-minute, 15-minute, and 1-hour intervals, and fed into a CNN-GRU model for training to learn the load variation characteristics at each time scale. The BMA method is then used to probabilistically fuse the prediction results of multiple CNN-GRU sub-models to construct an integrated prediction model. Finally, this model is used to output the predicted load power value. The technical effect is that through multi-timescale data collaboration and model integration, it effectively captures the macro-trends and micro-fluctuations of load, significantly improving prediction accuracy and robustness, and providing a reliable basis for energy dispatching in industrial parks.

[0098] like Figure 3 As shown, Figure 3 The diagram shows the network structure of the ensemble model (Convolutional Neural Network-Gated Recurrent Unit-Bayesian Model Averaging, CNN-GRU-BMA). In this model, multi-scale data refers to datasets with specific temporal granularity extracted from the original load sequence through different sampling intervals. The CNN-GRU hybrid model is a deep learning structure that integrates the local feature extraction capability of convolutional neural networks with the temporal modeling capability of gated recurrent units. Bayesian model averaging is a probabilistic fusion method that integrates the outputs of multiple models based on posterior probability weighting. Figure 3 The demonstrated CNN-GRU-BMA model operates as follows: First, data with sampling accuracies of 5 minutes (5T), 15 minutes (15T), and 1 hour (1H) are input into parallel-processed CNN-GRU sub-models. Each sub-model extracts local spatial features of the load sequence through convolutional layers, compresses the feature dimensions through pooling layers, and then feeds them into the GRU layer to capture long-term temporal dependencies. After each sub-model outputs preliminary prediction results, the Bayesian model averaging module calculates the posterior probability weights of each model based on the validation set performance, and performs probabilistic weighted fusion of the predicted values. The technical effect of this approach is that it simultaneously captures the micro-fluctuations and macro-trends of the load through multi-scale data collaborative analysis, leverages the advantages of hybrid models to enhance feature representation capabilities, and finally reduces the uncertainty of a single model through probabilistic fusion. Experimental results show that this method improves prediction accuracy by an average of over 0.85% and significantly enhances prediction robustness in complex scenarios such as extreme weather.

[0099] like Figure 4 As shown, Figure 4 In the diagram, T represents the initial time, S represents the fixed sliding step size, N represents the preset historical data length, W represents the feature sequence such as load power related data, P represents the prediction target sequence such as load power value, and the rolling prediction window refers to the dynamically updated data window. Figure 4The operating principle is as follows: starting from time T, the system trains the model from T to T+N based on historical window features W and the prediction target P, and predicts the future P value from T+N+1 to T+N+S. Then, the window slides for S steps to remove the earliest data from T to T+S and incorporate new data from T+N+S to T+N+2S to update the model and predict the next sequence. This iterative process achieves continuous prediction. Combined with other schemes in this application, the rolling prediction method is applied to each sub-model of the multi-timescale load power prediction sub-model. According to the sampling accuracy, a specific historical length N and prediction period S are set. The complete sequence of the target time period is generated by rolling and then fused by the average probability of the Bayesian model to finally output the fused prediction value. Therefore, it can continuously use the latest information to adapt to load time sequence changes, improve the long-term prediction accuracy and robustness, and provide reliable support for the economic dispatch of integrated energy parks.

[0100] like Figure 5 As shown, Figure 5 This paper compares the load power prediction performance of the CNN-GRU algorithm used in this application with other algorithms in a smart park. The horizontal axis represents the prediction time range from 00:00 on November 2nd to 08:00 on November 3rd, and the vertical axis represents the load power in kilowatts. The curve fluctuations reflect changes in the actual electricity demand of the park. Figure 5 The actual load power represents the real measured value as a comparison benchmark. The predicted values ​​of the CNN-GRU model, the Long Short-Term Memory (LSTM) model, and the CNN-LSTM model represent the prediction curves obtained by using different deep learning algorithms. The prediction time of 08:30 specifically refers to the model's prediction output for the future load at 8:30 am on November 2nd. Figure 5 The operating principle is to train the same historical time-series data using LSTM, CNN-LSTM, and the CNN-GRU hybrid model used in this invention, respectively. Based on the load power sequence for the next 32 hours obtained by the synchronous rolling prediction of each model, the predicted curves are visualized and compared with the actual curves. Figure 5 The data clearly shows that the prediction curve of the CNN-GRU model best matches the actual load power curve in terms of both shape and value. Especially during periods of rapid load change, the prediction error is significantly smaller than that of the LSTM and CNN-LSTM models. This verifies the advantages of the CNN-GRU hybrid structure in this application in collaboratively extracting local features and temporal dependencies. As the core unit of the multi-timescale prediction sub-model of this invention, it provides more reliable single-scale prediction results for subsequent Bayesian fusion, fundamentally improving the accuracy of the final integrated prediction.

[0101] like Figure 6 and Figure 7 As shown, Figure 6To compare the fluctuations of the model prediction evaluation metric (Root Mean Square Error, RMSE), this paper shows the fluctuations of the root mean square error of the three CNN-GRU sub-models and the CNN-GRU-BMA ensemble model at 5-minute, 15-minute, and 1-hour scales over 200 consecutive prediction time points. The horizontal axis represents the prediction time point number, and the vertical axis represents the RMSE value. Figure 7 At a specific prediction time, the actual load power is overlaid and compared with the curves of the sub-model predictions based on 5-minute, 15-minute, and 1-hour scale data and the predictions of the BMA integrated model as the future prediction point number changes. The horizontal axis is the future prediction point number, and the vertical axis is the load power in kilowatts.

[0102] Specifically, after obtaining historical multi-source data, this application can calculate the load power of the park using the following equation:

[0103] ;

[0104] in: It is the load power; This is the photovoltaic power generation capacity, which is generally positive during the power generation period. However, when the photovoltaic system is not generating electricity, it consumes a small portion of the electricity itself, resulting in a very small negative value. It refers to the power output of the network. It refers to internet access power. It represents the energy storage discharge power, which is positive. This is the energy storage charging power, which is positive.

[0105] Because load fluctuations are highly time-dependent, but the time characteristics in the data table cannot be parsed, they need to be converted into numerical features. A cosine transform is applied to the number of hours within a day, the number of days within a week, the number of days within a year, and the season to represent the periodic changes within that day, week, and year, including whether it is a weekday or a holiday. A feature column is added to distinguish between weekdays and rest days so that the deep learning model can learn more information. Taking the number of days within a week as an example, the feature calculation formula is as follows:

[0106] ;

[0107] in: It is the cosine transform value of the date within the week, and its value ranges between [-1, 1].

[0108] It is the weekday number of the predicted time, with a value of [0,6], such as 0 for Sunday and 1 for Monday;

[0109] This is the maximum value within the calculation period, which here is calculated based on the change in the number of days within the week. It lasts for 7 days.

[0110] The formula for the relative evaluation index RMSE of the prediction error of the calculation model in this invention is as follows:

[0111] ;

[0112] in: It is the total number of samples;

[0113] yes Actual load power at any given time;

[0114] yes Predicted load power at any given time;

[0115] The formula for calculating the accuracy of the load forecasting model is as follows: .

[0116] like Figure 8 As shown, an embodiment of this application provides a park load power prediction device 300, including a first acquisition module 302, a second acquisition module 304, a first determination module 306, a third acquisition module 308, a fourth acquisition module 310, a fifth acquisition module 312, and a first output module 314. The first acquisition module 302 is used to acquire historical multi-source data of the park, wherein the historical multi-source data includes at least load power related data; the second acquisition module 304 is used to acquire time-series data based on the historical multi-source data; the first determination module 306 is used to sample the time-series data based on multiple sampling accuracies to determine a multi-time-scale sampling dataset, wherein the multiple sampling accuracies are associated with multiple time scales; the third acquisition module 308 is used to... Based on the sampled dataset and the preset model, multiple load power prediction sub-models corresponding to multiple time scales are obtained. Each load power prediction sub-model is a hybrid model combining a convolutional neural network and a gated recurrent unit. The fourth acquisition module 310 is used to acquire the target time period. Based on the historical data length, single prediction period, and time series data corresponding to each load power prediction sub-model, the load power prediction sequence of each load power prediction sub-model within the target time period is acquired. The fifth acquisition module 312 is used to perform probability fusion based on multiple load power prediction sequences to acquire an integrated load power prediction model. The first output module 314 is used to output the fused load power prediction value corresponding to the target time period based on the integrated load power prediction model.

[0117] This application provides a park load power prediction device 300, including a first acquisition module 302, a second acquisition module 304, a first determination module 306, a third acquisition module 308, a fourth acquisition module 310, a fifth acquisition module 312, and a first output module 314. The first acquisition module 302 is used to acquire historical multi-source data of the park, wherein the historical multi-source data includes at least load power related data; the second acquisition module 304 is used to acquire time-series data based on the historical multi-source data; the first determination module 306 is used to sample the time-series data based on multiple sampling accuracies to determine a multi-time-scale sampling dataset, wherein the multiple sampling accuracies are associated with multiple time scales; the third acquisition module 308 is used to determine the sampling dataset based on the sampling data. Based on the data set and preset models, multiple load power prediction sub-models corresponding to multiple time scales are obtained. Each load power prediction sub-model is a hybrid model combining convolutional neural networks and gated recurrent units. The fourth acquisition module 310 is used to acquire the target time period. Based on the historical data length, single prediction period, and time series data corresponding to each load power prediction sub-model, the load power prediction sequence of each load power prediction sub-model within the target time period is acquired. The fifth acquisition module 312 is used to perform probabilistic fusion based on multiple load power prediction sequences to acquire an integrated load power prediction model. The first output module 314 is used to output the fused load power prediction value corresponding to the target time period based on the integrated load power prediction model.

[0118] like Figure 9 As shown in the illustration, an embodiment of this application provides a campus load power prediction device 400, including a processor 402 and a memory 404. The memory 404 stores programs or instructions. When the processor 402 executes the programs or instructions in the memory 404, it implements the steps of the campus load power prediction method as described in any of the above embodiments. Therefore, the campus load power prediction device possesses all the beneficial effects of the campus load power prediction method as described in any of the above embodiments.

[0119] This application provides a readable storage medium storing a program or instructions. When executed by a processor, the program or instructions implement the steps of the campus load power prediction method as described in any of the above embodiments. Therefore, the readable storage medium possesses all the beneficial effects of the campus load power prediction method as described in any of the above embodiments.

[0120] In one embodiment, the present application may further enhance the temporal features by using sine transform instead of cosine transform during the data preprocessing stage, and perform dimensionless processing by using Z-Score normalization instead of Min-Max normalization. In the model building stage, the activation function of the convolutional neural network can be ReLU, PReLU, or Sigmoid, and the gated recurrent unit can be replaced with RNN, LSTM, or other recurrent neural network variants. The number of layers and dimensionality parameters of the convolutional neural network and the gated recurrent unit can be dynamically adjusted according to the actual data characteristics. In the network optimization stage, the hyperparameter optimization method can use random search algorithm, Bayesian optimizer, or ant colony algorithm instead of particle swarm optimization. In the model ensemble stage, the Bayesian model averaging fusion method can be replaced with a stacking method. Integrated prediction is achieved by constructing a second-level learner with a BP neural network or convolutional neural network as the meta-model. At the same time, the sampling accuracy combination of the sub-models can be flexibly configured based on low-frequency, mid-frequency, and high-frequency scales. In terms of application scope, this application can be extended to multi-energy complementary scenarios such as photovoltaic power prediction, wind power prediction, and electricity market price prediction. By adjusting feature engineering and model parameters, it can adapt to the time series prediction needs of different energy systems.

[0121] In the claims, description, and accompanying drawings of this invention, the term "plural" refers to two or more. Unless otherwise explicitly defined, the terms "upper," "lower," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and simplifying the descriptive process, and are not intended to indicate or imply that the device or element referred to must have the described specific orientation, or be constructed and operated in a specific orientation. Therefore, these descriptions should not be construed as limiting the invention. The terms "connected," "installed," "fixed," etc., should be interpreted broadly. For example, "connected" can be a fixed connection between multiple objects, a detachable connection between multiple objects, or an integral connection; it can be a direct connection between multiple objects or an indirect connection between multiple objects through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in this invention can be understood based on the specific circumstances described above.

[0122] In the claims, description, and accompanying drawings of this invention, the terms "one embodiment," "some embodiments," "specific embodiment," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In the claims, description, and accompanying drawings of this invention, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0123] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the load power of a park, characterized in that, include: Acquire historical multi-source data of the park, wherein the historical multi-source data includes at least load power related data; Based on the aforementioned historical multi-source data, time-series data is obtained; The time-series data is sampled based on multiple sampling accuracies to determine a multi-time-scale sampling dataset, wherein the multiple sampling accuracies are associated with the multi-time-scale. Based on the sampled dataset and the preset model, multiple load power prediction sub-models corresponding to the multiple time scales are obtained, wherein each of the multiple load power prediction sub-models is a hybrid model combining a convolutional neural network and a gated recurrent unit. Obtain the target time period, and based on the historical data length and single prediction period corresponding to each load power prediction sub-model and the time series data, obtain the load power prediction sequence of each load power prediction sub-model within the target time period; A load power prediction ensemble model is obtained by probabilistically fusing multiple load power prediction sequences. Based on the load power prediction integration model, the load power fusion prediction value corresponding to the target time period is output.

2. The method for predicting park load power according to claim 1, characterized in that, The acquisition of time-series data based on the historical multi-source data includes: The historical multi-source data is cleaned, time-series aligned, and feature-engineered to obtain time-series data.

3. The method for predicting park load power according to claim 2, characterized in that, The process of cleaning, time-series alignment, and feature engineering of the historical multi-source data to obtain time-series data includes: For data with missing values ​​in the historical multi-source data, an interpolation method is used to imputate the missing values ​​to obtain preprocessed data. Specifically, for missing values ​​at the beginning and end of the data sequence, forward padding or backward padding methods are used for imputation. The preprocessed data with different sampling precisions are unified onto the same time index to obtain time series data.

4. The method for predicting park load power according to claim 1, characterized in that, The plurality of sampling accuracies include at least one low-frequency sampling accuracies, at least one medium-frequency sampling accuracies, and at least one high-frequency sampling accuracies.

5. The method for predicting park load power according to claim 4, characterized in that, The step of obtaining the load power prediction sequence for each of the load power prediction sub-models within the target time period includes: Based on the number of single prediction periods corresponding to each load power prediction sub-model, the target time period is divided to determine the number of rolling predictions required for prediction. Based on the length of the historical data, an initial historical data window is extracted from the time-series data; The first forward prediction is performed using the historical data window to obtain the first predicted sequence. The result of the first predicted sequence is used as the updated historical data to update the historical data window, and the historical data corresponding to the historical time furthest from the current time is deleted to keep the window length of the historical data window unchanged. Repeat the prediction and window update steps until the rolling prediction is completed for the specified number of times, and then concatenate each prediction sequence in order to determine the load power prediction sequence.

6. The method for predicting park load power according to claim 5, characterized in that, In each rolling forecast, the historical data window used for forecasting contains preprocessed real historical data, forecast data generated from the previous rolling forecast, or a combination of the real historical data and the forecast data generated from the previous rolling forecast.

7. The method for predicting park load power according to claim 4, characterized in that, The sampling time interval corresponding to the low-frequency sampling accuracy is 1 hour, the sampling time interval corresponding to the medium-frequency sampling accuracy is 15 minutes, and the sampling time interval corresponding to the high-frequency sampling accuracy is 5 minutes.

8. The method for predicting park load power according to any one of claims 1 to 7, characterized in that, The historical multi-source data also includes the grid connection power data of the park's grid connection points, energy storage charging and discharging power data, power generation power data, historical meteorological data of the park, meteorological forecast data, operating condition data, and production scheduling information of the park.

9. A power load prediction device for industrial parks, characterized in that, include: The first acquisition module is used to acquire historical multi-source data of the park, wherein the historical multi-source data includes at least load power related data; The second acquisition module is used to acquire time-series data based on the historical multi-source data; The first determining module is used to sample the time-series data based on multiple sampling accuracies to determine a multi-time-scale sampling dataset, wherein the multiple sampling accuracies are associated with the multiple time-scales; The third acquisition module is used to acquire multiple load power prediction sub-models corresponding to the multiple time scales based on the sampled dataset and the preset model, wherein each of the multiple load power prediction sub-models is a hybrid model combining a convolutional neural network and a gated recurrent unit. The fourth acquisition module is used to acquire the target time period and, based on the historical data length and single prediction period corresponding to each load power prediction sub-model and the time series data, acquire the load power prediction sequence of each load power prediction sub-model within the target time period. The fifth acquisition module is used to perform probabilistic fusion based on multiple load power prediction sequences to obtain an integrated load power prediction model; The first output module is used to output the load power fusion prediction value corresponding to the target time period based on the load power prediction fusion model.

10. A power load prediction device for industrial parks, characterized in that, include: processor; A memory containing programs or instructions, wherein the processor, when executing the programs or instructions in the memory, implements the steps of the park load power prediction method as described in any one of claims 1 to 8.

11. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the campus load power prediction method as described in any one of claims 1 to 8.

Images