Multi-element load prediction method and device for integrated energy system

Abstract

The invention provides a multi-element load prediction method and device for an integrated energy system, and belongs to the technical field of load prediction. The method provided by the invention comprises the steps of collecting multivariate load historical data and meteorological data, and generating a multivariate feature sequence according to the multivariate load historical data and the meteorological data; an Inform module and a BiGRU-GATT module are constructed, the multivariate feature sequence is input into the Inform module and the BiGRU-GATT module, the Inform module processes the multivariate feature sequence based on a probability sparse attention mechanism, and global features are output; the BiGRU-GATT module processes the multivariate feature sequence based on a global attention mechanism and outputs local features; and fusing the global features and the local features to obtain fused features, and outputting a prediction result based on the fused features. The multivariate load prediction method and device of the integrated energy system provided by the invention are used for improving the accuracy of multivariate load prediction.

Description

Technical Field

[0001] This application relates to the field of load forecasting technology, and in particular to a multi-source load forecasting method and apparatus for an integrated energy system. Background Technology

[0002] Multivariate load forecasting serves as a prerequisite for the optimized operation, scheduling decisions, and resource allocation of integrated energy systems. Its forecasting accuracy directly impacts the matching of energy supply, the economy of system operation, and stability. However, multivariate load data is influenced by a complex interplay of multiple factors, including meteorological conditions, user behavior patterns, and economic development levels. This results in significant characteristics such as strong nonlinearity, multi-timescale coupling, and high noise interference, leading to challenges in multivariate correlation modeling and insufficient cross-scale feature extraction, posing a severe challenge to accurate forecasting.

[0003] Current multivariate load forecasting methods are mainly divided into two categories: one is based on the extended application of traditional models, including classic algorithms such as Kalman filtering, time series analysis, and multiple linear regression, as well as traditional machine learning methods such as random forests, gradient boosting trees, and support vector machines. These methods largely follow the approach of single power load forecasting and are applied to multivariate load scenarios through simple adaptation. The other category is forecasting methods based on deep learning. With the development of deep learning technology, models such as convolutional neural networks, temporal convolutional networks, long short-term memory neural networks, gated recurrent units, and Transformers have been widely used in this field. Some studies have improved forecasting performance by optimizing algorithms or combining deep learning models. In recent years, to overcome the limitations of serial architectures, some studies have begun to try to use parallel architectures to achieve multi-model collaborative learning, in order to strengthen the complementary modeling of global and local information.

[0004] However, in existing technologies, traditional models and single machine learning methods do not specifically model the collaborative relationships between multiple loads, making it difficult to capture the complex coupling characteristics and dynamic dependencies between loads, and the prediction accuracy cannot meet the needs of integrated energy system scheduling. Mainstream deep learning models often neglect the accurate modeling of the time-series characteristics of multiple loads at multiple time scales, and when processing long-term series, they are prone to high computational complexity due to attention redundancy. Most combined deep learning models use a serial architecture for feature extraction and fusion, which has problems such as limited information transmission, low feature fusion efficiency, and insufficient coupling at multiple time scales. Some models are prone to error accumulation in multi-step prediction and have insufficient ability to capture local key features, resulting in poor prediction stability and accuracy in scenarios with large load fluctuations. Summary of the Invention

[0005] In view of this, this application provides a method and apparatus for predicting multiple loads in an integrated energy system, so as to improve the accuracy of multiple load prediction.

[0006] Specifically, this application is implemented through the following technical solution: The first aspect of this application provides a multi-source load forecasting method for an integrated energy system, the method comprising: Collect multi-dimensional load historical data and meteorological data of the integrated energy system, and generate multi-dimensional feature sequences based on the multi-dimensional load historical data and the meteorological data; An Informer module and a BiGRU-GATT module are constructed. The multivariate feature sequence is input into the Informer module and the BiGRU-GATT module respectively. The Informer module processes the multivariate feature sequence based on a probabilistic sparse attention mechanism and outputs global features. The BiGRU-GATT module processes the multivariate feature sequence based on a global attention mechanism and outputs local features. The global features and the local features are fused to obtain fused features, and the prediction result is output based on the fused features.

[0007] A second aspect of this application provides a multi-element load forecasting device for an integrated energy system, the device comprising an input module, a processing module, and an output module; wherein... The input module is used to collect multi-dimensional load historical data and meteorological data of the integrated energy system, and generate multi-dimensional feature sequences based on the multi-dimensional load historical data and the meteorological data. The processing module is used to construct an Informer module and a BiGRU-GATT module, and input the multivariate feature sequence into the Informer module and the BiGRU-GATT module respectively. The Informer module processes the multivariate feature sequence based on a probabilistic sparse attention mechanism and outputs global features; the BiGRU-GATT module processes the multivariate feature sequence based on a global attention mechanism and outputs local features. The output module is used to fuse the global features and the local features to obtain fused features, and output the prediction result based on the fused features.

[0008] The multi-variable load forecasting method and apparatus for integrated energy systems provided in this application focuses on the parallel operation architecture of the Informer module and the BiGRU-GATT module, breaking through the limitations of traditional single-model or serial architectures. The Informer module, based on a probabilistic sparse attention mechanism, reduces computational redundancy by selecting key vector pairs, while simultaneously extracting global features of long-term trends and cross-timescale correlations of multi-variable loads, solving the problems of attention redundancy and insufficient global feature capture in long-term series processing. The BiGRU-GATT module, relying on a global attention mechanism, focuses on dynamic fluctuations and local key features within short timescales, compensating for the inaccurate capture of local temporal dependencies in traditional models. The parallel design of the two modules enables the simultaneous extraction of global and local features, forming complementary advantages and providing core support for accurately modeling the complex relationships of multi-variable loads. By fusing the global features output by the Informer module and the local features output by the BiGRU-GATT module, deep integration of features at different time scales and dimensions is achieved. Based on the fused features, the prediction results are output, and the features extracted by the core module are transformed into accurate load prediction values. This effectively avoids the limitations of single feature modeling and ultimately achieves a significant improvement in the accuracy of multivariate load prediction, providing reliable decision support for the optimized scheduling of integrated energy systems. Attached Figure Description

[0009] Figure 1 A flowchart of an embodiment of the multi-element load forecasting method for integrated energy systems provided in this application; Figure 2 This is a schematic diagram of the structure of Embodiment 2 of the multi-element load forecasting device for the integrated energy system provided in this application. Detailed Implementation

[0010] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0011] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0012] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0013] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0014] Figure 1 This is a flowchart of an embodiment of the multi-element load forecasting method for the integrated energy system provided in this application. Please refer to... Figure 1 The method provided in this embodiment may include: S101. Collect multi-dimensional load historical data and meteorological data of the integrated energy system, and generate multi-dimensional feature sequences based on the multi-dimensional load historical data and the meteorological data.

[0015] Specifically, the process involves collecting historical data on multiple loads of the integrated energy system; collecting meteorological data related to these multiple loads; preprocessing the collected historical data on multiple loads and meteorological data; constructing feature categories based on the preprocessed historical data on multiple loads and meteorological data; and fusing the features of each category to generate a multi-feature sequence.

[0016] Furthermore, an integrated energy system is an integrated energy service system that integrates multiple energy forms such as electricity, heat, cooling, and natural gas, with the core objectives of efficient energy utilization and multi-energy coordinated supply. It coordinates all aspects of energy production, transmission, distribution, conversion, storage, and consumption. Historical data on diverse loads generated during the operation of the integrated energy system are collected. This data includes past operating records of the system's electricity load, cooling load, and heating load, ensuring that the data covers load changes under different time periods and operating conditions, providing fundamental data support for subsequent forecasts. Simultaneously, meteorological data closely related to changes in diverse loads, such as temperature, humidity, light intensity, and wind speed, are collected. Meteorological data is a crucial external variable affecting fluctuations in diverse loads, and its completeness and timeliness must be guaranteed. Preprocessing is performed on the two types of raw data. A minimization-maximization normalization method is used to process the preprocessed data, mapping it to a fixed interval to eliminate the influence of differences in data amplitude and avoid model training bias caused by inconsistent data scales.

[0017] Furthermore, after collecting historical load data and meteorological data from the integrated energy system, the method provided in this embodiment also includes: (1) Anomaly detection is performed on the multivariate load historical data and the meteorological data using box plots, and missing values ​​are identified based on the anomalies; Specifically, the process involves organizing historical multivariate load data and meteorological data, classifying and sorting them according to data type; drawing box plots for each category of data, and determining the quartiles, median, and outlier detection boundaries for each category based on the box plots; identifying outlier data points that exceed reasonable ranges based on outlier thresholds; distinguishing between genuine outliers and pseudo-outliers caused by missing data by combining data continuity and logical correlation; and marking the location of confirmed missing values ​​and their corresponding data categories.

[0018] Furthermore, the collected multivariate load historical data and meteorological data are categorized and organized according to their respective categories to ensure that the time dimension and format of each type of data are consistent, laying the foundation for subsequent box plot drawing. For example, the multivariate load historical data are divided into electrical load data, cooling load data, and heating load data, and the meteorological data are divided into multiple categories such as temperature, humidity, and light intensity. Box plots are drawn for each type of data. The box plots construct the data distribution range framework by displaying the minimum, first quartile, median, third quartile, and maximum value of the data. Among them, the interval between the first quartile and the third quartile is the interquartile range. Based on this, the outlier judgment boundary is set. Data that is less than the first quartile minus 1.5 times the interquartile range or greater than the third quartile plus 1.5 times the interquartile range is usually judged as outliers. This approach visually presents the dispersion of data and identifies anomalous data points that exceed the normal distribution range. Further analysis of these anomalous data points, considering the temporal continuity and logical correlation of the data, distinguishes between genuine outliers and pseudo-anomalies corresponding to missing values. For example, if the load data at a certain moment significantly deviates from the reasonable trend of adjacent moments and lacks corresponding objective reasons such as sudden meteorological changes, it is determined to be a genuine outlier. If the data at a certain moment exhibits obvious breaks or unreasonable gaps that cannot be explained by the changing patterns of adjacent data, it is determined to be a pseudo-anomaly caused by missing data, and this location is then marked as a missing value. The specific location and corresponding data category of all confirmed missing values ​​are recorded to form a missing value statistical list, providing a clear basis for subsequent data repair work.

[0019] (2) Repair the missing values ​​and normalize the repaired data.

[0020] Specifically, cubic spline interpolation was used to fill in and repair the missing values ​​of the labeled data; the units of measurement for the multivariate load data were standardized; the minimax normalization method was used to scale all the repaired data; and the rationality and effectiveness of the normalized data were verified.

[0021] Furthermore, cubic spline interpolation constructs a smooth cubic polynomial curve between adjacent valid data points containing missing values. Utilizing the curve's continuity and smoothness, it calculates reasonable data values ​​for the missing locations. This method effectively matches the changing trends of adjacent data, avoiding errors that may arise from simple linear interpolation. It ensures the consistency and continuity of the repaired data with the overall data sequence, effectively restoring the true changing patterns of the data. The repaired multivariate load data undergoes unified unit conversion. Since the original units of measurement for electricity, cooling, and heating loads may differ, to eliminate the impact of unit inconsistencies on subsequent model training and feature extraction, all load data are uniformly converted to power units, ensuring the comparability and consistency of different types of load data. All data after repair and unit unification are normalized using a minimax normalization method. By calculating the ratio of the difference between each data point and the maximum and minimum values ​​in the dataset, all data are mapped to a fixed interval of 0-1. For each data point, the minimum value of the dataset is subtracted, and then divided by the difference between the maximum and minimum values ​​of the dataset to obtain the normalized result. It can effectively eliminate the impact of amplitude differences between different data types, avoid the model from being overly biased towards data with large amplitudes during training, ensure that the contribution of various features to the model prediction is balanced, and improve the training efficiency and prediction accuracy of the model. It also verifies the rationality of the normalized data, checks whether the data are all within the 0-1 range, and whether there are any abnormal fluctuations or unreasonable values, to ensure the effectiveness of data processing and provide high-quality data support for the subsequent generation of multivariate feature sequences and model training.

[0022] By accurately detecting outliers and identifying missing values ​​in historical load and meteorological data using box plots, and then repairing missing values ​​using cubic spline interpolation and standardizing load data measurement units, maximum-minimum normalization is applied to effectively eliminate abnormal interference and fill data gaps to ensure data continuity and integrity. At the same time, it eliminates amplitude differences and unit barriers between different types of data, achieving unified standardization of data scale. This provides high-quality and reliable data support for subsequent multivariate feature sequence construction, multi-timescale feature extraction of models, and multivariate load coupling relationship modeling, helping to improve the accuracy and stability of multivariate load forecasting for integrated energy systems.

[0023] Furthermore, the steps for generating a multivariate feature sequence based on the multivariate historical load data and the meteorological data include: (1) Classify the multi-variable load historical data according to load type to obtain multiple load characteristics; Specifically, the load types included in the multi-source load historical data should be clearly defined; the multi-source load historical data should be classified according to the essential differences in the load types; and the data after each classification should be organized to form the corresponding load characteristics.

[0024] Furthermore, the collected historical data on diverse loads was analyzed to identify the core load types, typically including electricity load, cooling load, and heating load. Load types represent the main forms of energy consumption in a comprehensive energy system, each with its own unique patterns of change and influencing factors. Based on the essential differences in the attributes of each type of load, they were categorized as follows: electricity load mainly corresponds to the energy consumption data of various electrical equipment, such as industrial production electricity and residential electricity; cooling load mainly comes from the energy consumed by equipment such as air conditioners and refrigeration units to meet cooling needs; and heating load corresponds to the energy consumption data of heating equipment and water heaters to meet heating needs. Following this clear attribute definition, the historical data on diverse loads were categorized one by one, ensuring that each data point accurately belongs to its corresponding load type, avoiding confusion between different types of load data. The categorized load data was then systematically processed to remove any impurities that might have been introduced during the categorization process. The data sequence was then arranged chronologically to form a clear structure and well-defined multi-type load characteristics. Each type of load characteristic fully reflects the historical changes of its corresponding load type, providing accurate load dimension information for subsequent feature fusion.

[0025] (2) The meteorological data are classified according to the influence dimensions to obtain multiple types of meteorological characteristics; Specifically, the impact dimensions of meteorological data on multivariate load changes are identified; the meteorological data are classified according to the impact dimensions; and the data format of each classification is standardized to form multiple meteorological features.

[0026] Furthermore, based on the analysis of factors influencing multi-load changes, the key meteorological dimensions that play a crucial role in load fluctuations were identified. Through practical verification and theoretical analysis, temperature, humidity, light intensity, and wind speed were determined to be the core meteorological dimensions affecting multi-load changes. Temperature directly impacts the demand for cooling and heating loads, light intensity significantly affects the electricity load related to photovoltaic power supply, while humidity and wind speed indirectly affect various loads by influencing human comfort and equipment operating efficiency. According to these identified dimensions, the collected meteorological data were categorized: all data recording temperature information was classified as temperature characteristics, data recording air humidity as humidity characteristics, data reflecting light intensity as light intensity characteristics, and data describing airflow speed as wind speed characteristics. Other potentially secondary meteorological influencing factors were also categorized separately according to their respective dimensions to ensure no key meteorological information was overlooked. The meteorological data for each category was then formatted and standardized, unifying the data recording standards and timestamp formats to ensure consistency between the various meteorological data and load data in the time dimension, forming standardized multi-category meteorological features that facilitate subsequent collaborative analysis with other features.

[0027] (3) Extract the time information of the multivariate load historical data and the meteorological data, and classify multiple time features according to the time information; Specifically, time-related information is extracted from multivariate historical load data and meteorological data; the division dimensions of time features are determined; and the extracted time information is divided into multiple types of time features according to the division dimensions.

[0028] Furthermore, each record of the multivariate load historical data and meteorological data is traversed to extract time-related information, including year, month, day, hour, day of the week, and whether it is a holiday. Time information is an important basis for reflecting the temporal patterns of load changes; for example, load demand often differs significantly between different seasons, time periods, weekdays, and holidays. Based on the changing patterns of multivariate loads, the dimensions for classifying time features are determined. These dimensions mainly include time period dimensions, date dimensions, and special date dimensions. The time period dimension can be further subdivided into different time periods such as early morning, morning, afternoon, and evening; the date dimension can be divided into weekdays and weekends; and the special date dimension includes dates with special load change patterns, such as statutory holidays and traditional festivals. According to these classification dimensions, the extracted time information is categorized. Time information belonging to the same time period is grouped into corresponding time period features, time information of the same date type is grouped into corresponding date features, and time information of special dates is grouped into special date features. This classification method transforms abstract time information into multiple types of time features with clear physical meaning, effectively helping the model capture the changing patterns of load across different time dimensions.

[0029] (4) Integrate the various load characteristics, the various meteorological characteristics and the various time characteristics to generate a multivariate feature sequence.

[0030] Specifically, the time dimensions of multiple load characteristics, multiple meteorological characteristics, and multiple temporal characteristics are aligned; the three types of characteristics are integrated by feature splicing; the integrity and consistency of the integrated characteristics are verified, and a multivariate feature sequence is generated.

[0031] Furthermore, using timestamps as a benchmark, the previously obtained multiple load features, meteorological features, and time features are aligned along the time dimension. Each time node is checked one by one to ensure that at the same time point, each type of load feature, meteorological feature, and time feature has corresponding valid data. For data gaps at individual time points, the above data preprocessing methods are used to supplement and improve the data, ensuring the integrity and synchronization of the three types of features in the time series. The three types of features are then integrated using feature splicing. Using the time series as a guide, the various load features, meteorological features, and time features corresponding to the same time point are combined to form a feature vector containing multi-dimensional information. For example, the feature vector at a certain moment will simultaneously contain the power load, cooling load, and heating load data at that moment, meteorological data such as temperature, humidity, and light intensity, and time data such as the corresponding time period and date type, achieving deep fusion of features of different dimensions. The integrity and consistency of the integrated feature vector are then checked. It is checked whether each feature vector contains all necessary feature dimensions, whether the data format of various features is consistent, and whether there are obvious logical contradictions in the data. Problems found during the check are corrected in a timely manner. After verification, all feature vectors are arranged in chronological order to generate a complete multivariate feature sequence. This sequence comprehensively covers information from the three core dimensions of load, weather, and time, providing well-structured and complete input data for subsequent parallel feature extraction by the two modules.

[0032] By classifying historical load data according to load type to obtain multiple load features, and meteorological data according to impact dimension to obtain multiple meteorological features, the temporal information of the two types of data is extracted and divided into multiple temporal features. Then, the three types of features are integrated to generate a multi-dimensional feature sequence. This system can systematically sort out the core information related to multi-dimensional load forecasting, realize the structured decomposition and comprehensive integration of load, meteorological and temporal features. It not only retains the exclusive attributes of different types of features, but also builds a complete data system with interrelationships of multi-dimensional information. It provides high-quality input data with clear logic and complete information for subsequent parallel extraction of global and local features by dual modules, accurate modeling of multi-dimensional load coupling relationships and dynamic features at multiple time scales, and lays a solid data foundation for improving the accuracy of multi-dimensional load forecasting of integrated energy systems.

[0033] S102. Construct an Informer module and a BiGRU-GATT module. Input the multivariate feature sequence into the Informer module and the BiGRU-GATT module respectively. The Informer module processes the multivariate feature sequence based on a probabilistic sparse attention mechanism and outputs global features. The BiGRU-GATT module processes the multivariate feature sequence based on a global attention mechanism and outputs local features.

[0034] Specifically, an Informer module is constructed, which includes a probabilistic sparse self-attention mechanism, a self-attention distillation unit, and a generative decoder; a BiGRU-GATT module is constructed, which includes a bidirectional gated recurrent unit and a global attention mechanism; multivariate feature sequences are input into the Informer module, and key correlation information is filtered through the probabilistic sparse self-attention mechanism; the filtered information is processed by the self-attention distillation unit of the Informer module to enhance the long sequence feature representation; the global features are output through the generative decoder of the Informer module; multivariate feature sequences are input into the BiGRU-GATT module, and temporal features are extracted bidirectionally through the bidirectional gated recurrent unit; the features output by the bidirectional gated recurrent unit are weighted using the global attention mechanism; and the weighted local features are output.

[0035] Furthermore, an Informer module was constructed to extract global features. Its internal components include a probabilistic sparse self-attention mechanism, a self-attention distillation unit, and a generative decoder. These components work together to achieve efficient processing of long-term sequence data. Simultaneously, a BiGRU-GATT module was constructed. Focusing on local feature extraction, the BiGRU-GATT module consists of a bidirectional gated recurrent unit and a global attention mechanism, enabling it to accurately capture dynamic features and key information within short timescales.

[0036] Furthermore, the generated multivariate feature sequences are simultaneously input into the parallel Informer and BiGRU-GATT modules to achieve synchronous extraction of multidimensional features. The Informer module analyzes the correlation between each query vector and key vector in the feature sequence through a probabilistic sparse self-attention mechanism, estimates the sparsity of the query vectors by calculating the correlation divergence, and selects query-key pairs with important correlations for attention calculation, effectively reducing the interference of irrelevant information and reducing computational complexity while retaining key correlation features. Furthermore, the selected features are processed by a self-attention distillation unit, which successively reduces the sequence length, efficiently extracts the core information in long sequences, strengthens the feature expression of long-term dependencies, and uses a generative decoder to output the processed global features at once. The global features can reflect the overall changing trend of multivariate loads over long periods and the cross-period correlation rules, avoiding the error accumulation problem caused by stepwise decoding.

[0037] Furthermore, the BiGRU-GATT module processes the multivariate feature sequences through bidirectional gated recurrent units (BRUs). Each BRU contains independent forward and backward processing paths, capturing past and future information of the sequence respectively. The outputs of the forward and backward processing paths are fused through feature concatenation, fully preserving the temporal dependencies between the two paths while effectively mitigating the gradient vanishing problem and improving computational efficiency. A global attention mechanism is used to weight the fused features, comprehensively considering all hidden states output by the BRUs. Alignment weights are calculated based on the correlation between the current feature and each hidden state, and a weighted average is applied to all hidden states to generate local features. These local features accurately reflect the rapid fluctuations and local perturbations of multivariate loads within a short timescale, highlighting key time nodes.

[0038] Furthermore, the implementation steps for building the BiGRU-GATT module include: (1) Construct a bidirectional gated loop unit structure, and configure a forward unit and a reverse unit. The forward unit is used to capture the temporal dependency relationship of the multivariate feature sequence from the past to the present, and the reverse unit is used to capture the temporal dependency relationship of the multivariate feature sequence from the present to the future. Specifically, the overall framework of the bidirectional gated loop unit is constructed; the forward gated loop unit is configured to clarify its data processing direction; the reverse gated loop unit is configured to determine its data processing direction; and the core functions of the forward and reverse units are defined to correspond to the timing dependency capture requirements.

[0039] Furthermore, based on the basic structure of the gated loop unit, a bidirectional gated loop unit framework is constructed. This framework is a dual-path parallel structure, capable of processing multivariate feature sequences simultaneously from two opposite directions, overcoming the limitations of single-direction processing in capturing temporal information. A forward gated loop unit is configured, with its data processing direction set from the start time to the end time of the multivariate feature sequence, processing each feature data point from the past to the present in chronological order. The forward gated loop unit dynamically adjusts the degree of memory and forgetting of historical data through internal update and reset gates, accurately capturing the temporal dependencies in the multivariate feature sequence extending from the past to the present, such as the inheritance and change patterns between load data at different time periods. A reverse gated loop unit is also configured, with its data processing direction set from the end time to the start time of the multivariate feature sequence, processing feature data in reverse chronological order, from the present to the future. Similarly, by leveraging its internal gating mechanism, the reverse-gated recurrent unit can capture the temporal dependencies from the present to the future in a multivariate feature sequence, such as the potential impact of changes in meteorological factors in subsequent periods on the current load characteristics. Through the coordinated configuration of forward and reverse units, it achieves comprehensive coverage of the forward and backward temporal dependencies of the multivariate feature sequence, laying the foundation for subsequent feature extraction.

[0040] (2) Set the hyperparameters of the bidirectional gated loop unit and use the feature splicing method as the hidden layer output fusion method of the forward unit and the reverse unit; Specifically, the hyperparameter types of the bidirectional gated recurrent unit are determined; the specific values ​​of each hyperparameter are set according to the model training requirements and data characteristics; the advantages and disadvantages of different hidden layer output fusion methods are analyzed; the feature concatenation method is selected as the fusion method of the hidden layer outputs of the forward and backward units; and the specific implementation logic of feature concatenation is defined to ensure the integrity of the fused features.

[0041] Furthermore, the hyperparameters required for the operation of the bidirectional gated recurrent unit are analyzed. These hyperparameters mainly include the number of hidden layer neurons, training batch size, training epochs, dropout rate, and learning rate. These hyperparameters directly affect the unit's computational efficiency, feature extraction capability, and generalization performance. The specific values ​​of each hyperparameter are set based on the data scale and complexity of the multivariate feature sequences and the overall training objectives of the model. For example, an appropriate training batch size is determined based on the data sample size to avoid insufficient training memory due to an excessively large batch or instability due to an excessively small batch; the number of hidden layer neurons is adjusted according to the feature complexity to ensure that the unit has sufficient feature representation capability; a reasonable dropout rate is set to prevent model overfitting, and an appropriate learning rate is set to ensure convergence efficiency during training; common hidden layer output fusion methods are compared and analyzed, such as weighted summation, gated fusion, and feature concatenation. Weighted summation and gated fusion may result in the loss of some temporal features due to unreasonable weight allocation and may introduce additional trainable parameters, increasing model complexity and gradient interference risk. The feature concatenation method does not require the introduction of additional trainable parameters and can completely preserve the temporal features from the past to the present captured by the forward unit, as well as the temporal features from the present to the future captured by the reverse unit, avoiding the omission and distortion of feature information. The feature concatenation method was chosen as the fusion method, and its implementation logic is to directly concatenate the hidden layer outputs of the forward and reverse units according to the feature dimensions to form a fusion feature vector with higher dimensions and more comprehensive information, providing a rich temporal feature foundation for the subsequent global attention mechanism.

[0042] (3) After the bidirectional gated loop unit, a global attention mechanism layer is built, and the weight matrix and bias term for attention calculation are configured; Specifically, the basic architecture of the global attention mechanism layer is constructed; the weight matrix required for attention calculation is configured, and the matrix dimensions and initialization method are specified; the bias terms corresponding to attention calculation are configured, and their initial values ​​are set; the attention scoring function is defined, and the calculation logic of the correlation degree of temporal features is determined; the normalization method of attention weights is set to ensure the rationality of weight allocation.

[0043] Furthermore, the basic architecture of the global attention mechanism layer is built. The basic architecture takes the output features of the bidirectional gated recurrent unit as the processing object. The core function is to dynamically weight the temporal features, strengthen the expression of key features, and suppress the interference of irrelevant features. The weight matrix required for attention calculation is configured. The row and column dimensions of the weight matrix are determined according to the dimension of the output features of the bidirectional gated recurrent unit to ensure the compatibility of matrix operations. Initialize the weight matrix using appropriate methods such as random initialization or Xavier initialization to avoid gradient vanishing or exploding during model training due to excessively large or small initial values. Configure corresponding bias terms, typically setting the initial value of the bias terms to a zero vector or a small random value to adjust the attention calculation results and improve the model's fitting ability. Define an attention scoring function to calculate the correlation between the current target feature and all source features output by the bidirectional gated recurrent unit. Combining the characteristics of temporal features, select an appropriate scoring method and obtain a preliminary correlation score by calculating the similarity between the target feature and each source feature. Set a normalization method for the attention weights. Use the Softmax function to normalize the preliminary correlation score, converting the score into attention weights with values ​​between zero and one. The sum of the weights corresponding to all source features is one, ensuring the rationality of weight allocation and enabling the model to focus on key temporal features that have a significant impact on the prediction results.

[0044] (4) Integrate the bidirectional gated loop unit with the global attention mechanism layer to form the BiGRU-GATT module.

[0045] Specifically, the connection logic between the bidirectional gated loop unit and the global attention mechanism layer is determined; a data transmission channel between the bidirectional gated loop unit and the global attention mechanism layer is established; the structural integrity and logical rationality of the integrated module are verified; and the input and output interfaces of the BiGRU-GATT module are defined, and the data format requirements are clarified.

[0046] Furthermore, the connection logic between the bidirectional gated loop unit and the global attention mechanism layer is clarified. The fused temporal features output by the bidirectional gated loop unit are directly used as input data for the global attention mechanism layer, ensuring that the comprehensive temporal features captured by the bidirectional gated loop unit can be further optimized by the global attention mechanism layer. A data transmission channel is established between the bidirectional gated loop unit and the global attention mechanism layer, and the underlying logic of data flow is optimized to avoid delays or information loss during data transmission, ensuring that the output features of the bidirectional gated loop unit can be completely and efficiently transmitted to the global attention mechanism layer for processing. The integrity and logical rationality of the integrated module structure are verified. Check whether the input and output dimensions between the two layers match, whether the hyperparameter settings are consistent, and whether the gating mechanism and attention mechanism work smoothly. Correct any structural loopholes or logical contradictions that occur during the integration process in a timely manner. Define the input and output interfaces of the BiGRU-GATT module, clarify that the input data is a multivariate feature sequence, and the output data is a local feature vector after time-series feature extraction by bidirectional gated recurrent units and weighted enhancement by the global attention mechanism. Standardize the data format and dimensional requirements of the input and output to ensure that the BiGRU-GATT module can seamlessly connect with the Informer module and subsequent feature fusion modules to collaboratively complete the multivariate load prediction task.

[0047] By constructing a bidirectional gated recurrent unit structure containing forward and reverse units to comprehensively capture the forward and backward temporal dependencies of multivariate feature sequences, setting appropriate hyperparameters and using feature concatenation to fully preserve the output features of the bidirectional units, a global attention mechanism layer is subsequently built and corresponding weight matrices and bias terms are configured to enhance the expression of key features. Finally, the BiGRU-GATT module is integrated, achieving comprehensive capture of the temporal dependencies of multivariate feature sequences and precise enhancement of local key features. This effectively improves the module's ability to extract dynamic fluctuation features within a short timescale, while avoiding information loss and redundant parameter interference during feature fusion. It provides high-quality local feature support for subsequent global feature fusion with the Informer module and accurate prediction of multivariate loads, significantly enhancing the model's adaptability and modeling ability to complex dynamic changes in multivariate loads.

[0048] Furthermore, the Informer module processes the multivariate feature sequence based on a probabilistic sparse attention mechanism, and the steps for outputting global features include: (1) Perform vector transformation on the multivariate feature sequence to obtain the query vector matrix and the key vector matrix; Specifically, the feature dimensions and data length of the multivariate feature sequence are determined; vector transformation mapping rules are constructed to clarify the correspondence between the original features and vector elements; the multivariate feature sequence is converted into a query vector matrix according to the mapping rules; and the multivariate feature sequence is converted into a key vector matrix using the same mapping rules or adaptation rules.

[0049] Furthermore, the composition of the multivariate feature sequence is analyzed to clarify the number of feature dimensions it contains, such as the total dimension of features like load, weather, and time. Simultaneously, the time series length of the statistical data, i.e., the number of sample data points, is determined, providing a basis for vector transformation parameters. A vector transformation mapping rule is then constructed. Based on the attributes of the multivariate features, each time point's multivariate feature combination is considered a feature set. Each feature value in the set is mapped to a corresponding element of the vector in a fixed order, ensuring a one-to-one correspondence between the original feature information and the vector elements, without omitting any key feature information or introducing irrelevant redundant elements. According to the constructed mapping rule, the multivariate feature sequence is processed time-by-time, converting the feature set at each time point into a one-dimensional vector. All vectors corresponding to each time point are arranged in chronological order to form a query vector matrix. The number of rows in this matrix corresponds to the time series length, and the number of columns corresponds to the feature dimensions, completely preserving the temporal and feature information of the multivariate feature sequence. Using the same mapping rule as the query vector matrix, or an adapted mapping rule adjusted according to the needs of subsequent attention calculations, the same multivariate feature sequence is transformed again to generate a key vector matrix. Ensure that the key vector matrix and the query vector matrix are dimensionally compatible, meaning that the number of columns in the key vector matrix is ​​consistent with the number of columns in the query vector matrix, thus creating conditions for subsequent association calculations between the query vector and the key vector. Perform dimensional verification on the generated query vector matrix and key vector matrix, checking whether the number of rows and columns of the matrix meets the preset requirements and whether the dimensions between the two satisfy the operation rules for subsequent attention calculations. If there are any dimensional mismatch issues, adjust the mapping rules in a timely manner to correct them, ensuring that subsequent steps can be carried out smoothly.

[0050] (2) Calculate the sparsity of each query vector in the query vector matrix, filter target vector pairs from the query vector matrix and the key vector matrix based on the sparsity, and calculate the attention weight of the target vector pairs; Specifically, the calculation logic for query vector sparsity is defined; the sparsity of each query vector in the query vector matrix is ​​calculated one by one; a sparsity filtering threshold is set to clarify the criteria for judging target query vectors; based on the filtered target query vectors, the corresponding key vectors are matched to form target vector pairs; an attention weight calculation model is constructed to calculate the attention weight of each target vector pair; and the attention weights are normalized to ensure that the weight values ​​are reasonable.

[0051] Furthermore, the calculation logic for query vector sparsity is defined, and a core metric reflecting the degree of concentration of associations among elements in the vector is selected. By analyzing the association characteristics between query vectors and key vectors, the dispersion of the dot product between query vectors and key vectors is determined as the core metric for sparsity. This dispersion is quantified by calculating the correlation divergence; the larger the divergence value, the more concentrated the association between the vector and some key vectors, and the higher the sparsity. Sparsity is calculated for each query vector in the query vector matrix. For each query vector, its dot product with all key vectors in the key vector matrix is ​​calculated, and the sparsity value is obtained using a pre-defined divergence calculation formula. The sparsity results for each query vector are recorded. Considering the model's requirements for feature extraction accuracy and computational efficiency, a reasonable sparsity screening threshold is set. Query vectors with sparsity values ​​greater than the threshold are identified as target query vectors. Focusing on key related information is crucial for global feature extraction; otherwise, it is considered a non-target query vector and excluded. Based on the selected target query vectors, the corresponding key vectors are found in the key vector matrix. Each pair of target query vectors and their corresponding key vectors constitutes a target vector pair, ensuring that each target vector pair has strong correlation characteristics and can reflect the key temporal correlations in the multivariate feature sequence. Based on the dot product of the query vector and the key vector in the target vector pair, combined with parameters such as vector dimension, the initial attention weight of each target vector pair is calculated. The initial weight can reflect the importance of the vector pair in the global features. Then, an attention weight calculation model is constructed, and the initial attention weights of all target vector pairs are normalized. An appropriate normalization method is used to map the weight values ​​to a fixed interval, ensuring that the sum of all attention weights is one, making the weight allocation more reasonable and facilitating subsequent focusing on key features.

[0052] (3) Perform self-attention distillation on the attention weights to obtain the global features.

[0053] Specifically, a self-attention distillation framework is constructed; the convolutional layers, activation functions, and pooling operation parameters required for distillation are configured; attention weights are input into the distillation framework to perform one-dimensional convolution processing; the result after convolution is nonlinearly transformed by the activation function; pooling operations are used to filter and compress the transformed features; the distillation process is repeated to gradually strengthen the core features; and the final result after distillation, i.e., the global features, is output.

[0054] Furthermore, a self-attention distillation framework is constructed. The core objective of this framework is to purify and enhance the feature information contained in the attention weights, gradually reducing redundant information, highlighting key global features, and improving the efficiency of long sequence feature processing. The parameters of each component required for distillation are configured. Based on the dimension and feature complexity of the attention weight matrix, the kernel size and number of the one-dimensional convolutional layer are set to ensure that the convolutional layer can effectively extract local correlation information from the weight features. An appropriate non-linear activation function is selected to enhance the model's ability to fit complex features. The pooling window size and stride of the pooling operation are set to ensure efficient screening of key features and compression of feature dimensions. The normalized attention weights are input into the self-attention distillation framework, and one-dimensional convolution processing is performed. Through the sliding calculation of the convolution kernel on the attention weight matrix, local correlation patterns in the weight features are extracted, and information from adjacent weights is integrated to form more representative intermediate features. The intermediate features after convolution processing are input into a preset non-linear activation function to perform non-linear transformation on the features. This process breaks down the linear correlation between features, enhancing their nonlinear expressive power and enabling them to better reflect the complex changing patterns of multivariate loads. A defined pooling operation is used to process the nonlinearly transformed features, filtering key feature values ​​within each window through pooling windows. Simultaneously, the feature sequence length is compressed by a step size, achieving feature reduction and purification, reducing redundant information, and improving processing efficiency. Based on feature enhancement requirements, a distillation process of convolution processing-nonlinear transformation-pooling operation is repeatedly executed. Each round further purifies and enhances the features, gradually focusing on core global information such as long-term trends and cross-timescale correlations in the multivariate feature sequence. After multiple rounds of self-attention distillation, the final feature result, i.e., the global feature, is output. This feature accurately reflects the overall changing trend, cross-period correlation, and key global influencing factors of multivariate loads in long-term series, laying a solid foundation for subsequent fusion with local features.

[0055] By performing vector transformation on the multivariate feature sequence to obtain the query vector matrix and key vector matrix, the sparsity of each query vector is calculated to filter out key target vector pairs and calculate their attention weights. Self-attention distillation is then performed on the attention weights to obtain global features. This not only achieves accurate filtering of key correlation information in the multivariate feature sequence, effectively reducing interference from irrelevant information and computational complexity, but also gradually purifies and strengthens core features through self-attention distillation, efficiently capturing the overall changing trend and cross-period correlation patterns of multivariate loads in long-term series. The generated global features are both complete and representative, providing high-quality global dimensional support for subsequent fusion with local features and accurate prediction of multivariate loads, and significantly improving the model's ability to model long-term time-series dependencies.

[0056] Furthermore, the BiGRU-GATT module processes the multivariate feature sequence based on a global attention mechanism, and the steps for outputting local features include: (1) The BiGRU-GATT module captures the temporal dependencies of the multivariate feature sequences through forward and backward units respectively, and fuses the hidden layer outputs of the forward and backward units to output bidirectional temporal information; Specifically, the forward unit in the BiGRU-GATT module is activated to traverse the multivariate feature sequence in chronological order; the forward unit extracts the temporal dependencies of the multivariate feature sequence from the past to the present time step by time; the reverse unit in the module is activated to traverse the multivariate feature sequence in reverse chronological order; the reverse unit extracts the temporal dependencies of the multivariate feature sequence from the present to the future time step by time; the hidden layer outputs of the forward and reverse units are fused using a feature concatenation method; and the fused bidirectional temporal information is output.

[0057] Furthermore, the forward unit built into the BiGRU-GATT module is activated. The forward unit traverses the multivariate feature sequence sequentially from the start to the end. During this traversal, internal update and reset gates dynamically adjust the degree of memory and forgetting of historical feature information, accurately capturing the correlation between each moment's feature and past moment's features—that is, the temporal dependency from the past to the present, such as the impact of previous load changes on the current load state, or the inheritance relationship between historical meteorological data and current meteorological conditions. Simultaneously, the reverse unit in the module is activated. The reverse unit traverses the multivariate feature sequence in reverse chronological order, tracing back from the end to the start. Again, utilizing internal gate control mechanisms, the reverse unit focuses on capturing the potential correlation between each moment's feature and subsequent moment's features—that is, the temporal dependency from the present to the future, such as the prediction of future load demand based on current meteorological trends, or the impact of current energy consumption on future supply and demand balance. After the forward and reverse units complete their respective temporal dependency extraction, a feature concatenation fusion method is used to process the hidden layer outputs of both. This approach eliminates the need for additional trainable parameters, fully preserving both the forward temporal features captured by the forward unit and the backward temporal features captured by the backward unit, thus avoiding feature loss or interference from redundant parameters. By directly concatenating the hidden layer outputs of the two units along the feature dimension, a bidirectional temporal information containing complete forward and backward temporal correlation information is formed. This bidirectional temporal information is then used as input data for subsequent processing, providing a comprehensive foundation for the accurate generation of local features.

[0058] (2) Calculate the correlation between the predicted target and the bidirectional temporal information based on the global attention mechanism, and calculate the alignment weight vector based on the correlation; Specifically, the specific objectives of multivariate load forecasting are clarified, and the characteristic expression form of the forecasting objectives is determined; a correlation calculation model is constructed, and the correlation measurement standard between the forecasting objectives and bidirectional time series information is defined; based on the global attention mechanism, the correlation between the forecasting objectives and bidirectional time series information is calculated time by time; the correlation results at all times are collected and organized to form the original correlation sequence; and a normalization method is used to convert the original correlation sequence into an aligned weight vector.

[0059] Furthermore, the specific objectives of this multivariate load forecast are clearly defined, such as the future electricity load value, cooling load demand, or heating load scale for a specific period. These objectives are then transformed into corresponding feature expressions, enabling them to be correlated with bidirectional time-series information through dimensional matching. A scientific correlation calculation model is constructed based on a global attention mechanism. This model uses the feature expression of the forecast objective as a benchmark, defining a standard to measure the degree of correlation between the features at each moment in the bidirectional time-series information and the forecast objective. By calculating the similarity between the forecast objective features and the hidden states at each moment in the bidirectional time-series information, the correlation between the two is quantified; a higher similarity indicates a greater impact of the feature at that moment on the forecast objective. Following the constructed calculation model, the correlation between the forecast objective and the bidirectional time-series information is calculated moment by moment. All hidden states contained in the bidirectional time-series information are traversed, and the correlation score between each hidden state and the forecast objective features is calculated. The calculation results for each moment are fully recorded, forming the original correlation sequence. The original correlation sequence can intuitively reflect the differences in the importance of features at different times in the bidirectional time series information to the prediction target. A normalization method is used to transform the original correlation sequence, mapping all correlation scores to a value range of 0-1, with the sum of all scores being 1, ultimately forming an aligned weight vector. Normalization ensures the rationality of weight allocation, enabling the weight corresponding to each feature at each time point to accurately reflect its contribution to the prediction target, avoiding weight imbalance caused by differences in score scales.

[0060] Furthermore, the implementation steps for calculating the alignment weight vector based on the correlation include: 2.1 The softmax function is used to perform probability transformation on the correlation, mapping the correlation to weight coefficients; Specifically, the relevant dataset to be transformed is clearly defined, and the range and distribution characteristics of the data are confirmed; the softmax function is introduced as a probability transformation tool, and the calculation logic and application rules of the function are clarified; all relevant values ​​are substituted into the softmax function for value-by-value calculation; the range of values ​​of the weight coefficients after transformation is verified to ensure that they meet the probability distribution requirements.

[0061] Furthermore, all calculated correlation values ​​are collected to form a complete correlation dataset. The range of values ​​in this dataset is analyzed to understand the density and dispersion of the data distribution, providing a data foundation for subsequent probability transformation. The softmax function is introduced as a tool for converting correlation to weight coefficients. Each correlation value is converted into a positive number through exponential operations, and then each converted positive number is divided by the sum of all converted positive numbers, thereby mapping the original correlation values ​​to weight coefficients that conform to the probability distribution. Essentially, this process normalizes different correlation values, highlighting the weight proportion corresponding to high correlation values ​​while reasonably allocating the weight share of low correlation values. Each value in the correlation dataset is substituted into the softmax function for calculation. For each correlation value, its difference from other values ​​is first amplified through exponential operations, and then normalized to a value between 0 and 1, which is the weight coefficient of the corresponding hidden state. The magnitude of the weight coefficient directly reflects the closeness of the correlation between the corresponding hidden state and the prediction target; the larger the coefficient, the more significant the impact of the hidden state on the prediction result. The value range of all converted weight coefficients is verified to ensure that each coefficient is between 0 and 1, and the sum of all coefficients is 1, meeting the basic requirements of probability distribution. The rationality of the weight coefficients is analyzed, and in conjunction with the actual scenario of multivariate load forecasting, it is determined whether the hidden state corresponding to high correlation is indeed a key factor affecting the prediction target. If it is found that the weight coefficient allocation is inconsistent with the actual logic, the correlation calculation model or the application parameters of the softmax function are adjusted to ensure that the weight coefficients can truly reflect the importance of the hidden state.

[0062] 2.2 Construct an aligned weight vector based on the weight coefficients corresponding to all hidden states.

[0063] Specifically, determine the order of all hidden states and clarify their temporal relationships; arrange the corresponding weight coefficients sequentially according to the order of the hidden states; verify the completeness of the weight coefficient sequence to ensure that each hidden state has a corresponding weight coefficient; construct an aligned weight vector and clarify the dimensions and meaning of the elements of the vector; verify the effectiveness of the aligned weight vector to ensure that it can adapt to the subsequent weighted calculation requirements.

[0064] Furthermore, the arrangement order of all hidden states in the bidirectional time-series information is clearly defined, and this order is consistent with the time dimension of the multivariate feature sequence. That is, the hidden states are arranged according to the chronological logic from past to present to future, ensuring that the temporal attribute of each hidden state is clearly identifiable. Based on the determined arrangement order of the hidden states, the weight coefficients are arranged sequentially according to their correspondence. For example, the weight coefficient corresponding to the first hidden state is the first element of the vector, the weight coefficient corresponding to the second hidden state is the second element, and so on, forming an ordered sequence of weight coefficients. The completeness of the weight coefficient sequence is verified to check whether each hidden state has a unique corresponding weight coefficient, ensuring no omissions or duplicate matches. If a weight coefficient is missing or mismatched, the process is traced back to the correlation calculation or probability transformation steps to find the root cause and correct it, ensuring that the weight coefficient sequence completely corresponds to the set of hidden states. An aligned weight vector is constructed based on the ordered weight coefficient sequence. The dimension of this vector is consistent with the number of hidden states, and each element in the vector is the weight coefficient of the corresponding hidden state. The order of the vector is synchronized with the temporal order of the hidden states, intuitively reflecting the distribution of the influence of hidden states on the prediction target at different time points. Verify the effectiveness of the alignment weight vector by checking whether the vector dimension matches the size of the hidden state set and whether the order of the vector elements is consistent with the time dimension. Ensure that the vector can be directly applied to the weighted calculation of all subsequent hidden states, providing structured weight support for the accurate generation of local features.

[0065] By employing the softmax function to probabilistically transform the correlation between the prediction target and bidirectional temporal information, the original correlation values ​​are mapped to weight coefficients ranging from 0 to 1 with a total sum of 1. This approach highlights the significant proportion of highly correlated hidden states while rationally allocating weight shares for features with different degrees of correlation. By constructing aligned weight vectors according to the temporal order of the hidden states, a precise correspondence between weight coefficients and hidden states is achieved. This makes the weight allocation both logical and structured, effectively strengthening the contribution of key features to the prediction target while avoiding interference from irrelevant features. This provides accurate and standardized weight support for subsequent weighted calculations to generate features focusing on key local dynamics, thereby enhancing the targeting and recognizability of local features.

[0066] (3) Calculate the weighted value of all hidden states based on the alignment weight vector, and generate the local features based on the weighted value.

[0067] Specifically, all hidden states contained in the bidirectional time series information are extracted and arranged in chronological order to form a set of hidden states; the correspondence between the alignment weight vector and the set of hidden states is established, and the hidden state corresponding to each weight is identified; the hidden states are weighted one by one, and each hidden state is multiplied by its corresponding weight; all weighted hidden states are summed to obtain a comprehensive weighted value; the comprehensive weighted value is then integrated to generate local features that focus on key information.

[0068] Furthermore, from the bidirectional temporal information output by the BiGRU-GATT module, hidden states corresponding to all time steps are extracted. These hidden states contain temporal correlation details of multivariate feature sequences at different times. These hidden states are organized and sorted chronologically to form a structurally well-structured set, ensuring that the temporal attributes of each hidden state are clear and traceable. A one-to-one correspondence is established between the alignment weight vector and the hidden state set; that is, the weight value at each position in the alignment weight vector corresponds to a hidden state at the same time step in the hidden state set. This correspondence ensures that each hidden state receives a weight commensurate with its importance to the prediction target. Based on the established correspondence, each hidden state in the hidden state set is... The system performs a weighted calculation operation, multiplying the feature value of each hidden state by its corresponding weight value to obtain a weighted feature value for each hidden state. This strengthens the feature representation of key hidden states that have a significant impact on the prediction target, while weakening the influence of secondary hidden states. The weighted feature values ​​of all hidden states are then summed to integrate the scattered weighted feature values ​​into a comprehensive weighted value. This comprehensive weighted value centrally reflects the core features of all hidden states in the bidirectional time series information and highlights the contribution of features at critical moments. It can accurately reflect key information such as dynamic fluctuations and local perturbations of the multivariate feature sequence within a short time scale. Further feature integration is performed on the comprehensive weighted value to remove redundant information and extract core features, generating the final local feature. This local feature focuses on the rapid change patterns and key influencing factors of multivariate loads within a short period, complementing the global features output by the Informer module and providing high-quality local dimensional support for subsequent feature fusion and accurate multivariate load prediction. Simultaneously, the effectiveness of the local feature is verified to ensure that it accurately reflects the load change trend within a short time scale and meets the model's requirements for extracting local dynamic features.

[0069] By capturing the temporal dependencies of multivariate feature sequences bidirectionally through forward and backward units, and outputting complete bidirectional temporal information through feature splicing and fusion, the correlation between the prediction target and the bidirectional temporal information is calculated based on the global attention mechanism to obtain the alignment weight vector. Local features are generated by weighted calculation of all hidden states, which not only comprehensively covers the forward and backward temporal correlation information of multivariate feature sequences, but also accurately focuses on the key local dynamic features of the prediction target through attention weights. This effectively strengthens the expression of core information such as load fluctuations and local disturbances within a short time scale, avoids interference from irrelevant features, and the generated local features are highly targeted and distinctive. They form an efficient complement to the global features output by the Informer module, providing high-quality local dimension support for subsequent feature fusion and accurate prediction of multivariate loads, and significantly improving the model's ability to capture and model the complex dynamic changes of multivariate loads.

[0070] S103. The global features and the local features are fused to obtain fused features, and the prediction result is output based on the fused features.

[0071] Specifically, the system employs a feature concatenation method to fuse global and local features; it then uses a fully connected layer to reduce the dimensionality of the fused features; finally, it performs feature combination and nonlinear transformation on the reduced-dimensional features; outputs prediction results for multivariate loads; and evaluates the performance of the prediction results using preset evaluation metrics.

[0072] Furthermore, the global features output by the Informer module and the local features output by the BiGRU-GATT module are fused. The two types of features are integrated into a complete fused feature vector by feature concatenation. This can completely preserve all the information of global trend features and local dynamic features, realize the deep fusion of multi-time scale features and multi-dimensional correlation features, enhance the ability to express complex coupling relationships of multi-variable loads, and provide comprehensive feature support for accurate prediction.

[0073] The fused feature vector is input into a fully connected layer. The weight matrix of the fully connected layer performs a linear transformation on the fused features, achieving dimensionality reduction, eliminating redundant feature information, and extracting core fused features. An activation function is then used to perform a nonlinear transformation on the dimensionality-reduced core features, strengthening the nonlinear correlation between features and improving the model's ability to fit complex load variation patterns. Furthermore, feature combination further explores potential correlations between features of different dimensions, forming more representative predictive features. Based on the final processed features, the output layer of the fully connected layer outputs the prediction results for multi-variable loads, including predicted values ​​for electricity load, cooling load, and heating load.

[0074] Furthermore, in order to evaluate the accuracy and reliability of the prediction results, preset evaluation indicators are used to evaluate the performance of the prediction results, mainly including root mean square error, mean absolute percentage error, and weighted average absolute percentage error. By comparing the predicted values ​​with the actual load values, the prediction accuracy and stability of the model are comprehensively measured, providing a basis for the optimization and improvement of the model.

[0075] Furthermore, the steps for fusing the global features and the local features to obtain the fused features include: (1) Extract the first feature vector of the global feature and the second feature vector of the local feature respectively; Specifically, the core information dimensions of the global features are identified, and the extraction range of the first feature vector is determined; a transformation rule for the global feature vector is constructed; the first feature vector is extracted from the global features according to the transformation rule; the key dynamic dimensions of the local features are identified, and the extraction range of the second feature vector is determined; a transformation rule for the local feature vector is constructed, retaining the core dynamic information; the second feature vector is extracted from the local features according to the transformation rule.

[0076] Furthermore, the global features output by the Informer module are analyzed. These global features focus on long-term trends and cross-timescale correlations of diverse loads, clarifying their core information dimensions, such as long-term load fluctuation patterns and cross-period meteorological influences. Based on this, the extraction scope of the first feature vector is defined to ensure that the extraction process covers all key information dimensions of the global features. A conversion rule from global features to the first feature vector is constructed. According to the information structure of the global features, global information of different dimensions is mapped to vector elements in a fixed order, such as long-term trend features and cross-period correlation features. The quantification results of various types of information are mapped to different elements of the vector, ensuring the integrity of the global features and the one-to-one correspondence with the vector elements, without losing any key global information. Based on the constructed conversion rule, the global features are systematically processed and transformed into a one-dimensional first feature vector. The number of elements in this vector is consistent with the information dimensions of the global features, which can completely preserve the long-term time-series dependencies and overall change patterns contained in the global features.

[0077] Furthermore, the local features output by the BiGRU-GATT module are analyzed. The core of these features lies in key information such as load dynamic fluctuations and local disturbances within short timescales. Key dynamic dimensions are identified, such as short-term load mutation characteristics and local meteorological influence characteristics, to determine the extraction range of the second feature vector and ensure focus on core local dynamic information. A conversion rule from local features to the second feature vector is constructed. Referring to the conversion logic of global features, various local dynamic information is quantified and mapped into vector elements according to the information dimension order of the local features, focusing on retaining core information such as short-term fluctuations and key local node features, while avoiding redundant information interference. The local features are processed according to the conversion rule to generate a one-dimensional second feature vector. The number of elements matches the information dimension of the local features, accurately reflecting the load dynamic changes within short timescales. Dimensional verification is performed on the first and second feature vectors to ensure compatibility in vector structure, i.e., consistent vector arrangement and data type, creating conditions for subsequent dimensional concatenation operations. If dimensional incompatibility issues exist, the conversion rule is adjusted and corrected promptly.

[0078] (2) The first feature vector and the second feature vector are concatenated based on the dimensional direction of the feature vector to obtain the fused feature.

[0079] Specifically, determine the concatenation direction of the feature vectors as the dimensional direction and clarify the concatenation logic; align the concatenation references of the first and second feature vectors to ensure orderly concatenation; and concatenate the first and second feature vectors according to the dimensional direction.

[0080] Furthermore, the direction of feature concatenation is explicitly defined as the dimensional direction. This means that fusion is achieved by increasing the dimension of the feature vectors, rather than changing the time series length. This concatenation logic can simultaneously preserve the global information dimension of the first feature vector and the local information dimension of the second feature vector, achieving a comprehensive integration of the two types of features. The concatenation benchmark of the two feature vectors is aligned, using the starting position of the vectors as a unified benchmark to ensure that the first and second feature vectors are in the same order during concatenation, avoiding feature information confusion due to benchmark misalignment. For example, the concatenation benchmark can be set with the global feature dimension first and the local feature dimension last. The first and second feature vectors are concatenated according to the dimensional direction and the established benchmark. Specifically, all elements of the second feature vector are sequentially added to the end of the first feature vector to form a new high-dimensional vector, i.e., the fused feature. The dimension of this fused feature is the sum of the dimensions of the first and second feature vectors. It includes information such as the long-term trend and cross-timescale correlation of global features, and also integrates the short-term dynamic fluctuations and local key node features of local features, achieving a deep fusion of global and local information.

[0081] Furthermore, the dimensionality integrity of the merged features is verified. This involves checking whether the number of dimensions of the merged features equals the sum of the dimensions of the first and second feature vectors, and verifying that the feature information corresponding to each dimension is completely preserved without any missing or duplicated elements. If any dimensional anomalies are found, issues during the merging process are promptly investigated and corrected. By analyzing the information coverage of the merged features, it is confirmed whether they simultaneously encompass both global and local key information, and whether they can provide comprehensive feature support for subsequent multivariate load forecasting. This ensures that the merged features can reflect both the overall changing patterns of multivariate loads and capture dynamic fluctuations within short timescales, laying a solid feature foundation for improving forecast accuracy.

[0082] By extracting the first feature vector of the global features and the second feature vector of the local features respectively, and then concatenating them along the dimensional direction of the feature vectors, a fused feature is obtained. This not only fully preserves the core global information contained in the global features, such as the long-term changing trend of multi-loads and cross-time scale correlations, but also comprehensively integrates the core local information contained in the local features, such as short-term dynamic fluctuations and local key disturbances. This achieves deep complementary fusion of features at different time scales and in different dimensions, effectively enriching the information dimension and expressive power of the features, avoiding the limitations of a single feature dimension, and providing high-quality feature support with both integrity and detail for subsequent multi-load prediction. This significantly improves the modeling accuracy of the complex coupling relationship and dynamic change law of multi-loads.

[0083] Furthermore, the steps for outputting prediction results based on the fused features include: (1) Perform linear projection on the fused features to achieve low-dimensional spatial mapping of the fused features; Specifically, the high-dimensional composition of the fused features is analyzed to determine the number of feature dimensions to be reduced; a linear projection matrix is ​​constructed to determine the mapping relationship between the projected dimensions and the original dimensions; the operation rules of the linear projection are set to ensure the linear transformation characteristics of the projection process; the fused features are substituted into the linear projection matrix for operation to achieve dimensionality compression.

[0084] Furthermore, the dimensional composition of the fused features is analyzed. The fused features are composed of the first feature vector of the global features and the second feature vector of the local features. They have high dimensionality and contain multi-dimensional information. The total number of dimensions of the fused features is counted to clarify the dimensional distribution of various types of information such as global trends and local fluctuations, providing a dimensional basis for linear projection. Based on the needs of subsequent prediction tasks and the model's computational efficiency requirements, the target dimension in the low-dimensional space is determined to ensure that the target dimension can reduce computational complexity while fully preserving core feature information. Based on the original high-dimensional dimension and the target low-dimensional dimension, a linear projection matrix is ​​constructed. The number of rows in the matrix corresponds to the original feature dimension, and the number of columns corresponds to the target low-dimensional dimension. The values ​​of the matrix elements are determined through random initialization or orthogonal initialization to ensure the stability of the projection transformation.

[0085] Matrix multiplication is used to multiply high-dimensional fused features with a linear projection matrix. Through linear transformation, the high-dimensional features are mapped to a predefined low-dimensional space. Dimensional compression and information extraction are performed on the high-dimensional features to remove redundant correlations between features, retaining the core expressions of features from different dimensions, while reducing the computational load of subsequent model operations. The dimensions of the mapped low-dimensional features are validated to check if they match the predefined target dimensions. If a dimensional mismatch exists, the dimensional parameters of the linear projection matrix are adjusted and the calculation is recalculated. Simultaneously, the information integrity of the low-dimensional features is verified using metrics such as feature information entropy and variance to ensure that the projection process does not excessively lose key information such as global trends and local fluctuations. If excessive information loss occurs, the construction method of the projection matrix is ​​optimized or the target dimension is adjusted.

[0086] (2) Using a nonlinear activation function to perform a nonlinear transformation on the fused features mapped in the low-dimensional space, the multivariate load coupling relationship between the global features and the local features is determined; Specifically, a suitable nonlinear activation function is selected, and the nonlinear transformation characteristics of the function are clarified; the application parameters of the activation function are set to ensure that they are suitable for the data distribution of the low-dimensional fusion features; the fusion features after low-dimensional mapping are input into the nonlinear activation function for calculation; the expression form of the features after nonlinear transformation is analyzed to explore the potential correlation between features; based on the transformed features, the load correlation patterns corresponding to global features and local features are identified; the coupling strength and correlation law between multiple loads are quantified to determine the coupling relationship of multiple loads.

[0087] Furthermore, considering the strong nonlinearity of the multivariate load data, a suitable nonlinear activation function is selected. This function must possess good gradient propagation characteristics, effectively capturing the nonlinear correlations between features while avoiding gradient vanishing or exploding problems, ensuring the model can deeply explore complex feature relationships. Based on the data distribution range and numerical magnitude of the low-dimensional fused features, the application parameters of the activation function are set, such as adjusting the threshold and slope, so that the activation function can better adapt to the input features and fully utilize its nonlinear transformation capabilities. The fused features after low-dimensional space mapping are input element by element into the selected nonlinear activation function, and the features are transformed through the nonlinear operation of the function. This process breaks the linear constraints between features, strengthens the nonlinear correlation between global and local features, and makes the features better reflect the complex changing patterns of multi-load loads. It analyzes the numerical distribution and dimensional correlation of features after nonlinear transformation, and uses methods such as feature correlation analysis and cluster analysis to uncover potential correlation patterns between electricity, cooling, and heating loads corresponding to global and local features. Examples include the correlation between long-term temperature change trends and short-term cooling load fluctuations, and the complementary relationship between cross-period electricity load and heating load. Based on the uncovered correlation patterns, it quantifies the coupling strength between multi-load loads, clarifies the mutual influence patterns and synergistic change trends between different load types, such as the driving effect of electricity load growth on cooling load and the multi-load linkage patterns caused by meteorological factors, providing core correlation basis for subsequent load forecasting.

[0088] (3) Determine the load forecast results of the integrated energy system based on the multi-load coupling relationship.

[0089] Specifically, a load forecasting mapping model is constructed to clarify the corresponding logic between the coupling relationships of multiple loads and the forecast results; the output dimensions of the forecasting model are set to match the multiple load types of the integrated energy system; the determined coupling relationships of multiple loads are input into the forecasting mapping model for calculation; the rationality of the model output results is verified and judged in combination with the actual load change patterns; outliers in the forecast results are corrected to ensure the accuracy of the forecast results; and the final forecast results of the electricity, cooling, and heating loads of the integrated energy system are output.

[0090] Furthermore, based on the coupling characteristics of multiple loads, a scientific load forecasting mapping model is constructed. This model needs to establish a clear correspondence between coupling relationships and load forecast values. For example, quantitative indicators such as coupling strength and correlation patterns are used as inputs and mapped to corresponding load forecast values ​​through pre-defined calculation rules, ensuring the model fully utilizes the core information contained in the coupling relationships. According to the multiple load types of the integrated energy system, the output dimensions of the forecasting model are set, for example, the output dimensions correspond to electricity, cooling, and heating loads, ensuring the forecast results comprehensively cover the main load types of the system and meet scheduling optimization requirements. The coupling relationships of multiple loads, including quantitative information such as coupling strength and coordinated change patterns between loads, are input into the constructed forecasting mapping model for calculation. The model integrates key information in the coupling relationships and combines it with historical load change patterns to calculate preliminary results for future multiple load forecasts. These preliminary forecasts are then processed. The measurement results undergo rationality verification. This involves considering the actual operating scenarios of the integrated energy system, historical load data from the same period, current weather conditions, and user behavior patterns to determine if the prediction results conform to the objective laws of load changes. For example, it checks for anomalies such as load values ​​exceeding reasonable ranges or trends contradicting actual scenarios. For abnormal prediction values ​​discovered during verification, their causes are analyzed. If the anomaly is due to quantification deviations in coupling relationships, the process is backtracked to step two to correct the coupling relationship. If it's due to improper model parameter settings, the prediction mapping model parameters are adjusted, and the calculation is recalculated to correct the anomalies and improve prediction accuracy. The corrected prediction results for the integrated energy system's electricity, cooling, and heating loads are output, clearly defining the predicted values ​​and trends of various loads within a specific future period, providing accurate predictive support for the optimized scheduling and resource allocation of the integrated energy system.

[0091] By performing linear projection on the fused features to achieve low-dimensional spatial mapping, redundant information is effectively eliminated, computational complexity is reduced, and core features are retained. Then, a nonlinear activation function is used to perform nonlinear transformation on the low-dimensional features to accurately uncover the multi-load coupling relationship between global and local features. Finally, the load forecasting result of the integrated energy system is determined based on this coupling relationship. This approach not only achieves efficient optimization of feature dimensions but also fully captures the complex nonlinear correlations and dynamic changes of multi-loads, avoiding the limitations of single-feature modeling. It significantly improves the accuracy, stability, and reliability of load forecasting results, providing accurate and valuable decision support for the optimized scheduling and resource allocation of the integrated energy system.

[0092] The multi-variable load forecasting method for integrated energy systems provided in this embodiment, at the data level, uses box plots to detect outliers and identify missing values, combined with cubic spline interpolation repair and maximum-minimum normalization processing, effectively ensuring the integrity and consistency of historical multi-variable load data and meteorological data, laying a high-quality data foundation for subsequent modeling. By classifying and integrating three types of features—load, meteorology, and time—to generate multi-dimensional feature sequences, a structured fusion of multi-dimensional information is achieved, comprehensively covering key factors influencing load changes. At the feature extraction level, a parallel architecture of the Informer and BiGRU-GATT modules is adopted. The Informer module efficiently captures global features of long-term trends and cross-timescale correlations of multi-dimensional loads through probabilistic sparse attention mechanisms and self-attention distillation, significantly reducing computational complexity. The BiGRU-GATT module, relying on bidirectional gated recurrent units and global attention mechanisms, accurately captures dynamic fluctuations and local key features within short timescales, and strengthens the expression of core information by optimizing weight allocation through the softmax function. At the feature fusion and prediction level, global and local feature vectors are concatenated along the dimensional direction, achieving deep complementarity of features at different timescales. After linear projection dimensionality reduction and nonlinear activation transformation, the coupling relationships of multi-dimensional loads are accurately mined, effectively avoiding the limitations of single-feature modeling. The overall prediction error is lower than that of traditional models and single deep learning models, and it exhibits stronger stability and anti-error accumulation capabilities in load fluctuation scenarios. It can provide accurate and reliable decision support for the optimized scheduling and resource allocation of integrated energy systems, significantly improving the economy and stability of system operation.

[0093] Corresponding to the aforementioned embodiment of a multi-source load forecasting method for an integrated energy system, this application also provides an embodiment of a multi-source load forecasting device for an integrated energy system.

[0094] Figure 2 This is a schematic diagram of the structure of Embodiment 2 of the multi-element load forecasting device for the integrated energy system provided in this application. Please refer to... Figure 2 The apparatus provided in this embodiment includes an input module 210, a processing module 220, and an output module 230; wherein, The input module 210 is used to collect multi-dimensional load historical data and meteorological data of the integrated energy system, and generate a multi-dimensional feature sequence based on the multi-dimensional load historical data and the meteorological data. The processing module 220 is used to construct an Informer module and a BiGRU-GATT module, respectively inputting the multivariate feature sequence into the Informer module and the BiGRU-GATT module. The Informer module processes the multivariate feature sequence based on a probabilistic sparse attention mechanism and outputs global features; the BiGRU-GATT module processes the multivariate feature sequence based on a global attention mechanism and outputs local features. The output module 230 is used to fuse the global features and the local features to obtain fused features, and output prediction results based on the fused features.

[0095] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0096] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0097] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0098] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for multi-element load forecasting of an integrated energy system, characterized in that, The method comprises: collecting multi-element load historical data and meteorological data of the integrated energy system, and generating multi-element feature sequences according to the multi-element load historical data and the meteorological data; constructing an Informer module and a BiGRU-GATT module, inputting the multi-element feature sequences into the Informer module and the BiGRU-GATT module respectively, the Informer module processing the multi-element feature sequences based on a probabilistic sparse attention mechanism to output global features, and the BiGRU-GATT module processing the multi-element feature sequences based on a global attention mechanism to output local features; fusing the global features and the local features to obtain fused features, and outputting a prediction result based on the fused features.

2. The method of claim 1, wherein, The Informer module processes the multi-element feature sequences based on a probabilistic sparse attention mechanism to output global features, and comprises: performing vector conversion on the multi-element feature sequences to obtain a query vector matrix and a key vector matrix; calculating the sparsity of each query vector in the query vector matrix, screening a target vector pair from the query vector matrix and the key vector matrix based on the sparsity, and calculating the attention weight of the target vector pair; performing self-attention distillation processing on the attention weight to obtain the global features.

3. The method of claim 1, wherein, The BiGRU-GATT module processes the multi-element feature sequences based on a global attention mechanism to output local features, and comprises: The BiGRU-GATT module captures the time sequence dependency of the multi-element feature sequences through a forward unit and a reverse unit, fuses the hidden layer outputs of the forward unit and the reverse unit, and outputs bidirectional time sequence information; calculating the relevance of a prediction target and the bidirectional time sequence information based on the global attention mechanism, and calculating an alignment weight vector according to the relevance; calculating the weighted values of all hidden states according to the alignment weight vector, and generating the local features according to the weighted values.

4. The method of claim 1, wherein, The fusing of the global features and the local features to obtain fused features comprises: extracting a first feature vector of the global features and a second feature vector of the local features respectively; splicing the first feature vector and the second feature vector based on the dimension direction of the feature vectors to obtain the fused features.

5. The method of claim 1, wherein, The outputting of a prediction result based on the fused features comprises: performing linear projection on the fused features to realize low-dimensional space mapping of the fused features; performing nonlinear transformation on the fused features after low-dimensional space mapping by using a nonlinear activation function to determine the multi-element load coupling relationship between the global features and the local features; determining the load prediction result of the integrated energy system based on the multi-element load coupling relationship.

6. The method of claim 1, wherein, After collecting the multi-element load historical data and the meteorological data of the integrated energy system, the method further comprises: performing abnormal value detection on the multi-element load historical data and the meteorological data by using a box plot, and identifying missing values according to the abnormal values; repairing the missing values and performing normalization processing on the repaired data.

7. The method of claim 1, wherein, The generating the multi-element feature sequence according to the multi-element load historical data and the meteorological data comprises: classifying the multi-element load historical data according to load types to obtain multi-class load features; classifying the meteorological data according to influence dimensions to obtain multi-class meteorological features; extracting time information of the multi-element load historical data and the meteorological data, and dividing multi-class time features according to the time information; integrating the multi-class load features, the multi-class meteorological features and the multi-class time features to generate the multi-element feature sequence.

8. The method of claim 1, wherein, The constructing the BiGRU-GATT module comprises: building a bidirectional gated recurrent unit structure, configuring a forward unit and a reverse unit, the forward unit being used for capturing time sequence dependency of the multi-element feature sequence from the past to the present, and the reverse unit being used for capturing time sequence dependency of the multi-element feature sequence from the present to the future; setting hyperparameters of the bidirectional gated recurrent unit, and adopting a feature splicing manner as a hidden layer output fusion manner of the forward unit and the reverse unit; building a global attention mechanism layer after the bidirectional gated recurrent unit, and configuring a weight matrix and a bias term of attention calculation; integrating the bidirectional gated recurrent unit and the global attention mechanism layer to form the BiGRU-GATT module.

9. The method of claim 3, wherein, The calculating the alignment weight vector according to the correlation comprises: adopting a softmax function to perform probability conversion on the correlation, and mapping the correlation to a weight coefficient; constructing an alignment weight vector based on weight coefficients corresponding to all hidden states. 10.A device for multi-element load forecasting of an integrated energy system, characterized in that, The device comprises an input module, a processing module and an output module; wherein, the input module is used for collecting multi-element load historical data and meteorological data of the integrated energy system, and generating a multi-element feature sequence according to the multi-element load historical data and the meteorological data; the processing module is used for constructing an Informer module and a BiGRU-GATT module, inputting the multi-element feature sequence into the Informer module and the BiGRU-GATT module respectively, the Informer module processing the multi-element feature sequence based on a probability sparse attention mechanism to output global features, and the BiGRU-GATT module processing the multi-element feature sequence based on a global attention mechanism to output local features; the output module is used for fusing the global features and the local features to obtain fused features, and outputting a prediction result based on the fused features.

Images