Long-period power load prediction method and system based on double-model architecture

Abstract

The invention discloses a long-period power load prediction method and system based on a double-model architecture, and belongs to the technical field of power load prediction of a power system, and the method comprises the steps: obtaining related data of the power system, carrying out the preprocessing and feature extraction, and obtaining time sequence feature data; performing short-term probability prediction by using a first prediction model to obtain first quantile prediction data; performing long-term probability prediction by using a second prediction model to obtain second quantile prediction data; aligning prediction data, corresponding to a first preset time period, in the prediction data of the first quantile and the second quantile; calculating a probability distribution distance between the prediction data of the first quantile and the prediction data of the second quantile; and training the second prediction model by adopting a weighted combination loss function to obtain a long-period power load probability prediction result. According to the method, the problem of trend drift in long-period prediction is effectively improved, the calibration and practicability of the prediction interval are improved, and technical support is provided for risk management of a power system.

Description

Technical Field

[0001] This invention belongs to the field of power system load forecasting technology, specifically relating to a long-cycle power load forecasting method and system based on a dual-model architecture. Background Technology

[0002] Currently, in modern power systems, accurate load forecasting is a crucial prerequisite for power dispatch, market trading, and system planning. Power load sequences are influenced not only by historical loads but also by a combination of factors such as weather, calendar effects, economic activity, and renewable energy output, exhibiting significant nonlinearity, heteroscedasticity, and multi-periodicity.

[0003] As an improvement, current statistical methods (such as ARIMA and exponential smoothing) are effective in handling linear or near-linear trends, but their ability to model complex nonlinear and long-term dependencies is limited. In recent years, deep learning methods (such as CNN, RNN, LSTM, Transformers, and TimesNet) have made progress in end-to-end learning of complex time-series patterns, but still face two main challenges. For example, in long-term power load forecasting, information from short-term forecasts cannot be effectively transferred to longer forecasts, resulting in limited accuracy in long-term forecasts. Furthermore, the uncertainty information in forecasts cannot be effectively utilized during model training, leading to poor reliability of probabilistic forecasts. Summary of the Invention

[0004] This invention provides a long-cycle power load forecasting method and system based on a dual-model architecture. The purpose is to solve the problems in current long-cycle power load forecasting, such as the inability to effectively transfer information from short-term forecasts to longer forecasts, resulting in limited accuracy of long-cycle forecasts, and the inability to effectively utilize the uncertainty information of forecasts during model training, leading to poor reliability of probabilistic forecasts.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a long-cycle power load forecasting method based on a dual-model architecture, comprising the following steps: S1. Obtain relevant data of the power system, preprocess the data and extract features to obtain time series feature data for model input; S2. Based on time-series feature data, use the first prediction model to perform short-term probability prediction to obtain the first quantile prediction data corresponding to multiple time points within the first preset time period in the future. S3. Based on time-series feature data, a second prediction model is used to perform long-term probability prediction to obtain second quantile prediction data corresponding to multiple time points within a future second preset time period; the second preset time period includes the first preset time period. S4. Align the first quantile prediction data with the prediction data corresponding to the first preset time period in the second quantile prediction data; based on the aligned data, calculate the probability distribution distance between the first quantile prediction data and the corresponding second quantile prediction data; S5. The second prediction model is trained using a weighted combination loss function to obtain the long-cycle power load probability prediction results. The first prediction model is a probabilistic prediction model trained based on a gradient boosting decision tree framework, and the second prediction model is a multi-period probabilistic prediction model trained based on a temporal neural network architecture.

[0006] In some implementations, in S1, the power system-related data includes historical load data, meteorological data, and calendar information data; preprocessing includes outlier handling, missing value imputation, and data normalization; feature extraction includes extracting time period features, calendar flag features, and sliding window statistical features.

[0007] In some implementations, in S2, the first prediction model is the LightGBM model; the short-term probability prediction is a probability prediction of the load for the next day; and the first quantile prediction data is a matrix of prediction values ​​corresponding to multiple quantiles.

[0008] In some implementations, in S2, the loss function used to train the first prediction model is the quantile loss function Pinball Loss, as shown in the following formula: ; in, At the quantile level, For the loss of the LGB model, Number of time periods per day For the first time in a day Time period index For the first The predicted value at each time point, For the first The true value at any given moment.

[0009] In some implementations, in S3, the second prediction model is the TimesNet model; the long-term probability prediction is a probability prediction of the load over multiple future days; and the second quantile prediction data is a matrix of predicted values ​​corresponding to multiple quantiles.

[0010] Furthermore, in S3, the Times Net model contains multiple cascaded Times Block modules. Each Times Block module performs the following operations in sequence: performs a Fast Fourier Transform on the input sequence to analyze the spectrum and select the main frequencies; folds the one-dimensional time series into a two-dimensional tensor based on the period corresponding to the selected frequency; extracts features using a two-dimensional convolutional network; and reconstructs the extracted features back into one-dimensional space and performs weighted aggregation.

[0011] In some implementations, in S4, alignment specifically includes: aligning the first prediction model with the second... The prediction matrix for the predicted date and the corresponding value in the output matrix of the second prediction model. The submatrix of days, aligned by time points.

[0012] In some implementations, in S4, the method employs a two-stage training strategy: in the first stage, a first prediction model is trained independently using historical data; in the second stage, when training the second prediction model, the loss function simultaneously optimizes the quantile loss and the distribution similarity constraint with short-term predictions.

[0013] In some implementations, in S5, the loss function is a weighted sum of the main task loss and a constraint term based on the distance of the probability distribution. The expression for the weighted loss function is as follows: ; in, Loss to the main task For the second prediction model, the first Prediction of the celestial sphere For the first prediction model Prediction of the celestial sphere The distance is Wasserstein-1. For the weights of the constraint terms, For the true value, The prediction results output by the second prediction model. This represents the total loss of the dual-model architecture.

[0014] This invention also provides a long-cycle power load forecasting system based on a dual-model architecture to implement the aforementioned long-cycle power load forecasting method based on a dual-model architecture, including a time-series feature data module, a short-term probability forecasting module, a long-term probability forecasting module, a probability distribution distance calculation module, and a probability forecasting module; wherein: Time series feature data module: used to acquire power system related data, preprocess the data and extract features to obtain time series feature data for model input; Short-term probability prediction module: Used to perform short-term probability prediction based on time-series feature data and using the first prediction model to obtain the first quantile prediction data corresponding to multiple time points within the first preset time period in the future; Long-term probability prediction module: Used to perform long-term probability prediction based on time-series feature data and a second prediction model to obtain second quantile prediction data corresponding to multiple time points within a second preset time period in the future; the second preset time period includes the first preset time period; The probability distribution distance calculation module is used to align the first quantile prediction data with the prediction data corresponding to the first preset time period in the second quantile prediction data; based on the aligned data, it calculates the probability distribution distance between the first quantile prediction data and the corresponding second quantile prediction data. Probabilistic prediction module: used to train the second prediction model using a weighted combination loss function, and then obtain the long-term power load probabilistic prediction results; The first prediction model is a probabilistic prediction model trained based on a gradient boosting decision tree framework, and the second prediction model is a multi-period probabilistic prediction model trained based on a temporal neural network architecture.

[0015] Compared with existing technologies, the long-cycle power load forecasting method and system based on a dual-model architecture of the present invention has the following advantages: This invention presents a long-term power load forecasting method based on a dual-model architecture. By constructing a joint loss function that incorporates pinball loss and distribution similarity constraints, the two models are deeply integrated at the probabilistic prediction level, rather than simply stacking or weighted averaging models. This invention uses Wasserstein distance to measure the distribution similarity between short-term and long-term forecasts, thereby constraining and guiding the shape of the predicted distribution rather than just point estimation. By introducing the distribution information of LightGBM short-term forecasts as a soft constraint into the TimesNet model training process using Wasserstein distance, it achieves effective knowledge transfer across different forecast periods within the same task, effectively improving the trend drift problem in long-term forecasts. Furthermore, by combining quantile regression and distribution similarity constraints, it improves the calibration and practicality of the forecast interval, providing a relatively reliable basis for power system risk management and exhibiting good applicability. Attached Figure Description

[0016] The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0017] Figure 1 This is a flowchart illustrating an embodiment of a long-cycle power load forecasting method based on a dual-model architecture according to the present invention. Figure 2 This is a schematic diagram of the knowledge transfer mechanism in a long-cycle power load forecasting method based on a dual-model architecture according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to indicate that selected embodiments of the invention are based on the embodiments of the invention, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the invention.

[0020] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0021] It should be noted that the apparatus and methods disclosed in the embodiments herein can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments herein. In this regard, each block in a flowchart or block diagram may represent a module, program, or part of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system to perform the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0022] In addition, the functional modules in the various embodiments of this article can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0023] How can we provide a long-term power load forecasting method that can transfer high-quality information learned from near-future short-term forecasts to longer-term forecasts to improve long-term accuracy? How can we effectively utilize forecast uncertainty information during model training to improve the calibration and reliability of probabilistic forecasts?

[0024] like Figure 1 As shown, the present invention provides a long-cycle power load forecasting method based on a dual-model architecture, comprising the following steps: S1. Obtain relevant data of the power system, preprocess the data and extract features to obtain time series feature data for model input; S2. Based on time-series feature data, use the first prediction model to perform short-term probability prediction to obtain the first quantile prediction data corresponding to multiple time points within the first preset time period in the future. S3. Based on time-series feature data, a second prediction model is used to perform long-term probability prediction to obtain second quantile prediction data corresponding to multiple time points within a future second preset time period; the second preset time period includes the first preset time period. S4. Align the first quantile prediction data with the prediction data corresponding to the first preset time period in the second quantile prediction data; based on the aligned data, calculate the probability distribution distance between the first quantile prediction data and the corresponding second quantile prediction data; S5. The second prediction model is trained using a weighted combination loss function to obtain the long-cycle power load probability prediction results. The first prediction model is a probabilistic prediction model trained based on a gradient boosting decision tree framework, and the second prediction model is a multi-period probabilistic prediction model trained based on a temporal neural network architecture.

[0025] This invention presents a long-term power load forecasting method based on a dual-model architecture. It employs a forecasting approach that combines a gradient-boosting decision tree-based model with a time-series neural network-based model, and introduces a knowledge transfer mechanism based on probability distribution distance. This mechanism aligns the output of the short-term model with the corresponding output of the long-term model in the same time period, calculates the probability distribution distance between them, and integrates this distance as a constraint term into the loss function of the long-term model. This achieves directed knowledge transfer from short-term to long-term forecasts. Leveraging the higher accuracy of the short-term model in near-future forecasts, it constrains and guides the long-term model's forecast distribution within the same time interval, effectively improving the accuracy of long-term point and trend forecasts. This invention integrates probabilistic forecasting, allowing knowledge transfer to simultaneously affect the central trend and uncertainty range of the forecast values. By making the output distribution of the long-term model within the short-term forecast interval closer to the reliable distribution of the short-term model, it improves the rationality and calibration of the long-term model's uncertainty estimation throughout the entire forecast interval, providing a more reliable quantitative basis for power system risk management.

[0026] Furthermore, the first and second prediction models of this invention employ LightGBM and TimesNet, respectively. In training both models, a weighted loss function incorporating Wasserstein-1 distance is introduced to ensure knowledge transfer. Moreover, this invention guarantees strict comparability of short-term and long-term prediction information across the time dimension through data alignment.

[0027] As an alternative, the two-stage training strategy of this invention makes the model training process more stable and efficient. The two-stage training strategy first ensures the performance of the short-term predictor, and then uses its high-quality output to optimize the long-term predictor, which can further improve the reliability of prediction.

[0028] This invention also provides a long-cycle power load forecasting system based on a dual-model architecture, including a time-series characteristic data module, a short-term probability forecasting module, a long-term probability forecasting module, a probability distribution distance calculation module, and a probability forecasting module; wherein: Time series feature data module: used to acquire power system related data, preprocess the data and extract features to obtain time series feature data for model input; Short-term probability prediction module: Used to perform short-term probability prediction based on time-series feature data and using the first prediction model to obtain the first quantile prediction data corresponding to multiple time points within the first preset time period in the future; Long-term probability prediction module: Used to perform long-term probability prediction based on time-series feature data and a second prediction model to obtain second quantile prediction data corresponding to multiple time points within a second preset time period in the future; the second preset time period includes the first preset time period; The probability distribution distance calculation module is used to align the first quantile prediction data with the prediction data corresponding to the first preset time period in the second quantile prediction data; based on the aligned data, it calculates the probability distribution distance between the first quantile prediction data and the corresponding second quantile prediction data. Probabilistic prediction module: used to train the second prediction model using a weighted combination loss function, and then obtain the long-term power load probabilistic prediction results; The first prediction model is a probabilistic prediction model trained based on a gradient boosting decision tree framework, and the second prediction model is a multi-period probabilistic prediction model trained based on a temporal neural network architecture.

[0029] The following detailed description of a long-cycle power load forecasting method and system based on a dual-model architecture, using specific embodiments, further illustrates the present invention.

[0030] This invention employs a dual-model architecture, including: using LightGBM for short-term prediction, leveraging its advantages in efficiently handling feature interactions and calendar effects to output a probability prediction for a single day (D+1) (96 points / day × 5 quantiles); and using TimesNet for long-term prediction, utilizing its multi-period modeling capabilities to output a 960-point × 5 quantile matrix for multiple days (D+1~D+10). The 5 quantiles are as follows: .

[0031] like Figure 2 As shown, this invention first strictly aligns the outputs of LightGBM and TimesNet according to time. The D+1 day quantile prediction matrix output by LightGBM is then aligned with the corresponding day quantile prediction matrix output by TimesNet, item by item, according to time and five quantile points, to obtain a matrix with a consistent shape for subsequent similarity constraints and loss calculations. For example, the LightGBM matrix corresponding to the... The 96×5 of the predicted date corresponds to the TimesNet output matrix. The 96×5 pairs are aligned, and then the "Wasserstein-1 distance" is calculated on the aligned 96×5 pairs and added to the training loss of TimesNet as a constraint.

[0032] This invention employs a weighted combination of training losses for TimesNet: ; in Indicates TimesNet for the first Quantile prediction for the day (the first day) The ( )th in the 960×5 matrix 1)· 96: (96 lines) For the corresponding LightGBM output; The main task loss (in this invention, the loss is the loss on the actual observed values) (quantile pinball loss) The similarity distance between the prediction results of the two models (Wasserstein-1 distance in this invention) is the distance between the prediction results of the two models. The weights of the constraint terms are used to balance prediction accuracy and trend consistency. For the true value, The prediction results output by the second prediction model. This represents the total loss of the dual-model architecture.

[0033] Specifically: such as Figure 1 As shown, the present invention provides a long-cycle power load forecasting method based on a dual-model architecture, which is carried out through the following steps: Step 1: Data preparation and collection. Integrate historical load data, meteorological data (temperature, wind speed, weather conditions), and calendar information (holidays, seasonal markers). Standardize the meteorological data, such as extracting temperature values ​​and assigning ordered codes to weather conditions (e.g., sunny: 1, cloudy: 2, overcast: 3, etc.). Temporal feature extraction: Extracting multidimensional features from timestamps, including: Time cycle characteristics: month, day, hour, day of the year, week of the year; Calendar indicators: whether it is the end of the month, the beginning / end of the season, the weekend, or the beginning of the month; Sliding window statistical features: lag features and window statistical features (mean, maximum, minimum, etc.).

[0034] Data preprocessing: Remove outliers, such as extreme load values ​​caused by equipment malfunction or data logging errors. Handle missing data, which can be filled using interpolation or by averaging adjacent data points.

[0035] Data normalization: Employ the min-max normalization method to ensure that the data falls within a certain range. Interval: Step 2: LightGBM model training.

[0036] LightGBM was trained using rolling window data, and probability predictions were output using quantile regression, yielding 5 quantiles: ; The loss function for the predicted values ​​at 96 time points (15-minute intervals) in the next day (D+1) is Pinball Loss.

[0037] ; in At the quantile level, For the loss of the LGB model, This refers to the number of time slots per day, where the time slot interval is 15 minutes. =96, For the first time in a day Time period index For the first Forecast values ​​(forecast load) at each time point. For the first The actual value (actual load) at each moment.

[0038] Step 3: TimesNet model construction.

[0039] The TimesNet time series model is used to capture load variation patterns in different periods such as daily, weekly, and monthly using its multi-period modeling capabilities.

[0040] The model structure is composed of multiple stacked TimesBlocks connected by residuals. Each TimesBlock includes: Fast Fourier Transform analysis of the spectrum, selecting the k frequencies with the largest amplitude; folding the one-dimensional time series into a two-dimensional tensor based on the selected period; extracting temporal features using a 2D convolutional network (Inception block); reconstructing the features back into one-dimensional space and weighting and aggregating them; and outputting a prediction matrix of 960 time points × 5 quantiles for the next 10 days (D+1~D+10).

[0041] Step 4: Design of knowledge transfer mechanism and loss function.

[0042] Alignment strategy: The LightGBM prediction matrix (96×5) for day i is strictly aligned with the corresponding day i submatrix (96×5) in the TimesNet output matrix according to the time point.

[0043] The training loss for TimesNet is weighted and combined: ; in Indicates TimesNet for the first Quantile prediction for the day (the first day) The ( )th in the 960×5 matrix 1)· 96: (96 lines) For the corresponding LightGBM output; For the quantile loss of the main TimeNet task, The asserstein-1 distance between the predictions of the two models. The weights of the constraint terms are used to balance prediction accuracy and trend consistency.

[0044] Step 5: Model training.

[0045] Two-stage training strategy: LightGBM model pre-training: The LightGBM model is trained independently using historical data to optimize quantile prediction accuracy.

[0046] TimesNet model training: Initialize TimesNet parameters and simultaneously optimize quantile loss and distribution similarity constraints with short-term predictions during training.

[0047] In summary, this invention presents a long-cycle power load forecasting method and system based on a dual-model architecture. The method, based on LightGBM-Timesnet and probabilistic forecasting, first trains the LightGBM model as a short-term quantile forecaster, then transfers or uses its quantile forecasting information on day D+1 as a constraint to the Timesnet model for long-cycle forecasting. By incorporating quantile-based curve similarity constraints into the Timesnet model training, the Timesnet model is guided to learn a more reasonable trend and uncertainty structure, thereby improving the performance of long-cycle forecasting and the quality of uncertainty estimation, which has certain practical significance.

[0048] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Anyone skilled in the art can readily implement the present invention according to the description and above. Any modifications, alterations, or equivalent variations made using the technical content disclosed above are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. A long-term power load forecasting method based on a dual-model architecture, characterized by, The method comprises the following steps: S1, obtaining power system related data, preprocessing and feature extraction of the data, and obtaining time series feature data for model input; S2, based on the time series feature data, using a first prediction model for short-term probability prediction, and obtaining first quantile point prediction data corresponding to multiple time points in a future first preset period; S3, based on the time series feature data, using a second prediction model for long-term probability prediction, and obtaining second quantile point prediction data corresponding to multiple time points in a future second preset period; The second preset period includes the first preset period; S4, aligning the first quantile point prediction data with the prediction data corresponding to the first preset period in the second quantile point prediction data; Based on the aligned data, the probability distribution distance between the first quantile point prediction data and the corresponding second quantile point prediction data is calculated; S5, training the second prediction model using a weighted combination loss function, and then obtaining a long-period power load probability prediction result; The first prediction model is a probability prediction model trained based on a gradient boosting decision tree framework, and the second prediction model is a multi-period probability prediction model trained based on a time series neural network architecture. 2.The long-term power load forecasting method based on the dual-model architecture according to claim 1, wherein, In the S1, the power system related data includes historical load data, weather data and calendar information data; the preprocessing includes outlier processing, missing value filling and data normalization; the feature extraction includes extracting time period features, calendar marker features and sliding window statistical features. 3.The long-term power load forecasting method based on the dual-model architecture of claim 1, wherein, In the S2, the first prediction model is a LightGBM model; the short-term probability prediction is a probability prediction of the load in the future one day; and the first quantile point prediction data is a prediction value matrix corresponding to multiple quantile points. 4.The long-term power load forecasting method based on the dual-model architecture of claim 1, wherein, In the S2, the loss function used for training the first prediction model is a quantile loss function PinballLoss as follows: ； wherein, is the quantile level, is the loss of the LGB model, is the number of daily time periods, is the index of the th time period within the day, is the prediction value for the th time instant, is the true value for the th time instant. 5.The long-term power load forecasting method based on the dual-model architecture of claim 1, wherein, In the S3, the second prediction model is a TimesNet model; the long-term probability prediction is a probability prediction of the load in the future multiple days; and the second quantile point prediction data is a prediction value matrix corresponding to multiple quantile points. 6.The long-term power load forecasting method based on the dual-model architecture according to claim 5, wherein, In the S3, the Times Net model includes multiple cascaded Times Block modules, and each Times Block module sequentially performs the following operations: performing fast Fourier transform on the input sequence to analyze the frequency spectrum and selecting the main frequency; folding the one-dimensional time sequence into a two-dimensional tensor based on the period corresponding to the selected frequency; extracting features using a two-dimensional convolution network; and reconstructing the extracted features back to one-dimensional space and performing weighted aggregation. 7.The long-term power load forecasting method based on the dual-model architecture of claim 1, wherein, In the S4, the aligning specifically comprises: aligning, by time point, a prediction matrix of the first prediction model for a to-be-predicted day with a corresponding sub-matrix of the second prediction model output matrix for the day. day.​ 8.The long-term power load forecasting method based on the dual-model architecture of claim 1, wherein, In the S4, the method adopts a two-stage training strategy: in the first stage, the first prediction model is independently trained using historical data; and in the second stage, when training the second prediction model, the loss function simultaneously optimizes the quantile loss and the distribution similarity constraint with the short-term prediction. 9.The long-term power load forecasting method based on the dual-model architecture of claim 1, wherein, In the S5, the loss function is composed of a main task loss and a constraint term based on the probability distribution distance, and the expression of the weighted combination loss function is as follows: ； wherein, is the main task loss, is the quantile prediction of the second prediction model for day , is the quantile prediction of the first prediction model for day , is the Wasserstein-1 distance, is the constraint term weight, is the true value, is the prediction result output by the second prediction model, is the total loss of the dual model architecture.

10. A long-term power load forecasting system based on a dual-model architecture for implementing the long-term power load forecasting method based on a dual-model architecture according to any one of claims 1-9, characterized in that, The method comprises a time series feature data module, a short-term probability prediction module, a long-term probability prediction module, a probability distribution distance calculation module and a probability prediction module; wherein: The time series feature data module is configured to acquire power system related data, pre-process and extract features of the data, and obtain time series feature data for model input. The short-term probability prediction module is configured to perform short-term probability prediction based on the time series feature data using a first prediction model, and obtain first quantile point prediction data corresponding to multiple time points within a first preset time period in the future. The long-term probability prediction module is configured to perform long-term probability prediction based on the time series feature data using a second prediction model, and obtain second quantile point prediction data corresponding to multiple time points within a second preset time period in the future; the second preset time period includes the first preset time period. The probability distribution distance calculation module is configured to align the first quantile point prediction data with prediction data corresponding to the first preset time period in the second quantile point prediction data; and calculate a probability distribution distance between the first quantile point prediction data and the corresponding second quantile point prediction data based on the aligned data. The probability prediction module is configured to train the second prediction model using a weighted combination loss function, and thereby obtain a long-period power load probability prediction result. The first prediction model is a probability prediction model trained based on a gradient boosting decision tree framework, and the second prediction model is a multi-period probability prediction model trained based on a time series neural network architecture.

Images