A multi-element load prediction method for an integrated energy system and related devices

Abstract

The application relates to the technical field of load prediction, in particular to a comprehensive energy system multi-element load prediction method and related devices. The method comprises the following steps: acquiring original data of a comprehensive energy system; obtaining a comprehensive data set through the original data; determining a plurality of candidate features; determining N target features, taking the target features as inputs of a load prediction model, combining outputs of the load prediction model, adopting a JSSAO algorithm to iteratively train the load prediction model, and obtaining optimal hyperparameters; inputting the target features into the load prediction model configured with the optimal hyperparameters, obtaining enhanced features, processing the enhanced features, and obtaining a time sequence feature vector; and inputting the time sequence feature vector into a load decoder to obtain a multi-element load prediction result. The application can provide a high-precision and strong-adaptability solution for comprehensive energy system multi-element 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 related apparatus for integrated energy systems. Background Technology

[0002] Load forecasting is a core support for the safe and efficient operation of energy systems, and its accuracy directly affects the rationality of energy dispatch, the controllability of operating costs, and the stability of supply and demand balance. As the global energy structure transitions towards cleaner and more diversified energy sources, integrated energy systems, as key carriers integrating multiple energy forms such as electricity, cooling, and heating, play an irreplaceable role in improving energy utilization efficiency, promoting renewable energy consumption, and reducing carbon emissions. Accurate forecasting of multiple loads (electricity, cooling, and heating loads) is a prerequisite for integrated energy systems to achieve optimized dispatch and efficient resource allocation—only by accurately grasping the temporal variation patterns of different loads and their coupling relationships can operational strategies that balance economy and reliability be formulated to avoid energy waste or supply shortages. However, existing multi-load joint forecasting models still have significant limitations when facing the strong coupling, high volatility, and characteristic differences of electricity, cooling, and heating loads in user-level integrated energy systems. For example, single deep learning architectures (basic Transformer, shared decoder LSTM, etc.) lack sufficient adaptability to the differences in multi-load characteristics and insufficient guarantee of temporal causality. Summary of the Invention

[0003] To address the aforementioned issues, this application provides a method and related apparatus for predicting multiple loads in an integrated energy system, which offers a highly accurate and adaptable solution for predicting multiple loads in an integrated energy system.

[0004] According to one aspect of the embodiments of this application, a multi-element load forecasting method for integrated energy systems is proposed, the method comprising: Acquire raw data from the integrated energy system, including electrical load, heat load, cooling load, and meteorological data; The electrical load, heat load, cooling load, and meteorological data are fused to obtain a comprehensive dataset. The comprehensive dataset includes load sequence characteristics for four seasons, and the load sequence characteristics for each season include electrical load sequence characteristics, heat load sequence characteristics, and cooling load sequence characteristics. Multiple candidate features are identified in the comprehensive dataset, and these candidate features are those associated with the electrical load, the thermal load, and the cooling load. Among the candidate features, N target features with the highest correlation to the electrical load, the thermal load, and the cooling load are identified, where N is a positive integer greater than or equal to 1; The target features are used as input to a preset load forecasting model, and combined with the output of the load forecasting model, the JSSAO algorithm is used to iteratively train the load forecasting model to obtain the optimal hyperparameters for the load forecasting model output by the JSSAO algorithm. The target features are input into the load prediction model configured with the optimal hyperparameters, and enhanced features are obtained through forward dilation convolution and backward reverse dilation. The enhanced features are processed by the masked multi-head attention layer of the load prediction model to obtain a temporal feature vector; The time-series feature vector is input into the load decoder of the load prediction model to obtain the multivariate load prediction results related to the electrical load, the heat load, and the cooling load output by the load decoder.

[0005] In the above scheme, the meteorological data includes temperature data, air pressure data, and relative humidity data. The fusion of the electrical load, heat load, cooling load, and meteorological data to obtain a comprehensive dataset includes: The electrical load, thermal load, cooling load, temperature data, air pressure data, and relative humidity data are fused to obtain fused data. The fused data is divided into four seasons to obtain sub-data corresponding to the four seasons; For each season, the sub-data corresponding to the season is decomposed into time series to obtain the electrical load sequence characteristics, heat load sequence characteristics, and cold load sequence characteristics corresponding to the season. The electrical load sequence characteristics, heat load sequence characteristics, and cold load sequence characteristics corresponding to the season are then fused with the temperature data, air pressure data, and relative humidity data corresponding to the season to obtain the fused characteristics corresponding to the season. The fusion features corresponding to each season are fused to obtain the comprehensive dataset.

[0006] In the above scheme, determining the N target features with the highest correlation to the electrical load, the heat load, and the cooling load from among the candidate features includes: For each candidate feature, a comprehensive correlation score is determined relating the candidate feature to the electrical load, the thermal load, and the cooling load. The top N candidate features with the highest comprehensive correlation scores among all the candidate features are selected as the target features.

[0007] In the above scheme, the step of using the target feature as input to a preset load forecasting model, and combining it with the output of the load forecasting model, to iteratively train the load forecasting model using the JSSAO algorithm to obtain the optimal hyperparameters for the load forecasting model output by the JSSAO algorithm includes: The population of hyperparameters for the load prediction model is initialized by Latin hypercube sampling to obtain an initial population, which includes multiple individuals, each representing a set of hyperparameter combinations. The model prediction value for each individual is determined by the load prediction model, and the fitness score for each individual is determined based on the model prediction value for each individual. The initial population is divided into an exploratory subgroup and a development subgroup. The exploratory subgroup focuses on global search, while the development subgroup focuses on local fine-grained search. The exploration subgroup and the development subgroup are iteratively trained, and the population division ratio of the exploration subgroup and the development subgroup is dynamically adjusted with the number of iterations; After a preset number of iterations of training, the individual with the highest fitness score in the merged subgroup formed by the exploration subgroup and the development subgroup is selected as the optimal individual. The combination of hyperparameters corresponding to the optimal individual is taken as the optimal hyperparameters.

[0008] In the above scheme, the step of inputting the target features into the load prediction model configured with the optimal hyperparameters, and obtaining enhanced features through forward dilation convolution and backward reverse dilation, includes: The input sequence corresponding to the target feature is subjected to forward dilation convolution to obtain the forward dilation convolution result, so as to capture the continuous change trend of local time-dependent learning load; The input sequence corresponding to the target feature is subjected to backward reverse expansion to obtain the backward expansion convolution result, so as to capture the recovery trend after the peak of reverse local dependency learning; The enhanced feature is obtained by adding the target feature to the forward dilation convolution result and the backward dilation convolution result.

[0009] In the above scheme, the step of processing the enhanced features through the masked multi-head attention layer of the load prediction model to obtain the temporal feature vector includes: The input sequences corresponding to the enhanced features are respectively mapped to a query matrix Q, a key matrix K, and a value matrix V; The enhanced features are processed by the query matrix Q, the key matrix K, and the value matrix V to obtain the temporal feature vector.

[0010] In the above scheme, the load decoder includes an electrical load decoder, a thermal load decoder, and a cold load decoder. The step of inputting the time-series feature vector into the load decoder of the load prediction model to obtain the multivariate load prediction results output by the load decoder related to the electrical load, the thermal load, and the cold load includes: The time-series feature vectors are respectively input into the electrical load decoder, the thermal load decoder, and the cold load decoder to obtain the electrical load prediction result output by the electrical load decoder, the thermal load prediction result output by the thermal load decoder, and the cold load prediction result output by the cold load decoder. The multi-load forecast result is determined based on the electrical load forecast result, the heat load forecast result, and the cooling load forecast result.

[0011] According to one aspect of the embodiments of this application, a multi-element load forecasting device for an integrated energy system is proposed, the device comprising: The acquisition unit is used to acquire raw data of the integrated energy system, including electrical load, heat load, cooling load and meteorological data; The fusion unit is used to fuse the electrical load, the heat load, the cooling load, and the meteorological data to obtain a comprehensive dataset. The comprehensive dataset includes load sequence characteristics for four seasons, and the load sequence characteristics for each season include electrical load sequence characteristics, heat load sequence characteristics, and cooling load sequence characteristics. The first determining unit is configured to determine multiple candidate features in the comprehensive dataset, wherein the candidate features are features associated with the electrical load, the thermal load, and the cooling load; The second determining unit is used to determine, among the candidate features, the N target features that are most relevant to the electrical load, the thermal load and the cooling load, where N is a positive integer greater than or equal to 1; The training unit is used to take the target features as input to a preset load prediction model, and combine them with the output of the load prediction model to iteratively train the load prediction model using the JSSAO algorithm to obtain the optimal hyperparameters for the load prediction model output by the JSSAO algorithm. The input unit is used to input the target features into the load prediction model configured with the optimal hyperparameters, and obtain enhanced features through forward dilation convolution and backward reverse dilation; An enhancement unit is used to process the enhancement features through the masked multi-head attention layer of the load prediction model to obtain a temporal feature vector; The prediction unit is used to input the time-series feature vector into the load decoder of the load prediction model to obtain the multivariate load prediction results related to the electrical load, the heat load, and the cooling load output by the load decoder.

[0012] According to one aspect of the embodiments of this application, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the multi-variable load forecasting method for integrated energy systems as described above. According to one aspect of the embodiments of this application, a computer program product is provided, the computer program product including a computer program that is read and executed by a processor of an electronic device, causing the electronic device to perform the integrated energy system multi-load forecasting method as described above.

[0013] The beneficial effects of this application are as follows: This application integrates the electrical load, heat load, cooling load, and meteorological data to obtain a comprehensive dataset composed of load sequence features from four seasons. Multiple candidate features are identified within this comprehensive dataset, and the N target features with the highest correlation to the electrical load, heat load, and cooling load are then determined. These target features are used as input to a pre-defined load prediction model, and combined with the output of the load prediction model, the JSSAO algorithm is used to iteratively train the load prediction model. This yields the optimal hyperparameters for the load prediction model output by the JSSAO algorithm, thereby improving the prediction accuracy of the load prediction model.

[0014] In the internal architecture of the load forecasting model, the target features are input into the load forecasting model configured with the optimal hyperparameters. Enhanced features are obtained through forward dilation convolution and backward reverse dilation. The enhanced features are then processed by the masked multi-head attention layer of the load forecasting model to obtain a temporal feature vector. Based on this, the temporal feature vector is input into the load decoder of the load forecasting model to obtain the multivariate load forecasting results related to the electrical load, the heat load, and the cold load output by the load decoder.

[0015] This application addresses the significant differences in the characteristics (driving factors, fluctuation patterns, and time patterns) of electricity, cooling, and heating loads throughout the four seasons by designing a seasonally adaptable single-load independent modeling strategy, thus solving the problem of insufficient adaptability of multi-task shared decoders in scenarios with differentiated characteristics. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of the integrated energy system multi-element load forecasting method provided in the embodiments of this application; Figure 2A flowchart illustrating the multi-source load forecasting method for integrated energy systems provided in this application embodiment; Figure 3 This is the overall logic diagram of multi-source load forecasting for an integrated energy system provided in the embodiments of this application; Figure 4 This is a flowchart of the JASSO algorithm proposed in this application; Figure 5 This is the architecture diagram of the traditional MMHAtt; Figure 6 This is the architecture diagram of RBiTC-MMHAtt proposed in this application; Figure 7 A block diagram of a multi-load forecasting device for an integrated energy system provided in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of a terminal provided in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of a server provided in an embodiment of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the solutions of this application, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] It should be noted that while some processes described in the specification, claims, and accompanying drawings include multiple steps appearing in a specific order, it should be clearly understood that these steps may not be performed in the order they appear herein, or may be performed in parallel. The step numbers are merely used to distinguish different steps and do not themselves represent any execution order. Furthermore, descriptions such as "first," "second," or "objective" in this document are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. "Multiple" in this document refers to at least two.

[0019] It is worth noting that in the specific embodiments of this application, raw data, meteorological data, and other related data of the integrated energy system are involved. When the above embodiments of this application are applied to specific products or technologies, permission or consent from the target object is required, and the collection, use, and processing of related data must comply with relevant laws, regulations, and standards. For example, when the embodiments of this application need to obtain raw data, meteorological data, and other related data of the integrated energy system, separate permission or consent from the target object can be obtained through pop-up windows or redirection to a confirmation page. After obtaining the separate permission or consent from the target object, the raw data, meteorological data, and other related data of the integrated energy system used to enable the embodiments of this application to operate normally can then be obtained.

[0020] Please see Figure 1 , Figure 1 This is a system architecture diagram of the integrated energy system multi-source load forecasting method provided in this application embodiment. It includes a terminal 140, an Internet connection 130, a gateway 120, a server 110, etc.

[0021] Terminal 140 can take various forms, including desktop computers, laptops, PDAs (personal digital assistants), mobile phones, vehicle terminals, and dedicated terminals. Furthermore, it can be a single device or a collection of multiple devices. For example, multiple desktop computers can be interconnected via a local area network, sharing a single monitor to work collaboratively, forming a single terminal 140. Terminal 140 can communicate with the Internet 130 via wired or wireless means to exchange data.

[0022] Server 110 refers to a computer system capable of providing certain services to terminal 140. Compared to ordinary terminal 140, server 110 has higher requirements in terms of stability, security, and performance. Server 110 can be a single high-performance computer in a network platform, a cluster of multiple high-performance computers, a portion of a single high-performance computer (e.g., a virtual machine), or a combination of portions of multiple high-performance computers (e.g., virtual machines). Server 110 can also communicate with the Internet 130 via wired or wireless means to exchange data.

[0023] Gateway 120, also known as an internetwork connector or protocol converter, is a computer system or device that acts as a translator, enabling network interconnection at the transport layer. It bridges the gap between two systems using different communication protocols, data formats, languages, or even completely different architectures. Gateways can also provide filtering and security functions. Messages sent from terminal 140 to server 110 are forwarded to the corresponding server 110 via gateway 120. Messages sent from server 110 to terminal 140 are also forwarded to the corresponding terminal 140 via gateway 120.

[0024] Load forecasting is a core support for the safe and efficient operation of energy systems. Its accuracy directly affects the rationality of energy dispatch, the controllability of operating costs, and the stability of supply and demand balance. As the global energy structure transitions towards cleaner and more diversified energy sources, integrated energy systems, as key carriers integrating multiple energy forms such as electricity, cooling, and heating, play an irreplaceable role in improving energy utilization efficiency, promoting renewable energy consumption, and reducing carbon emissions. Accurate forecasting of multiple loads (electricity, cooling, and heating loads) is a prerequisite for integrated energy systems to achieve optimized dispatch and efficient resource allocation. Only by accurately grasping the temporal variation patterns of different loads and their coupling relationships can operational strategies that balance economy and reliability be formulated, avoiding energy waste or supply shortages. Furthermore, multi-load forecasting is also of great significance to the power system: as a core component of integrated energy systems, the load fluctuations of the power system directly affect grid frequency stability and voltage control. The strong correlation between cooling, heating, and electrical loads (e.g., air conditioning electricity drives cooling loads, and electric heating affects heating loads) enables multi-load collaborative forecasting to provide more comprehensive decision-making basis for peak shaving and valley filling and renewable energy consumption in the power system, contributing to the safe and stable operation of the new power system. However, existing multi-load joint forecasting models still have significant limitations when facing the strong coupling, high volatility, and characteristic differences of electricity, cooling, and heating loads in User Integrated Energy Systems (UIES): Traditional joint forecasting models (such as multi-task learning MTL hard-shared architecture) struggle to balance the global coupling relationships (such as electricity-heat linkage) and local detailed features (such as short-term peak electricity consumption) of multiple loads, often focusing on shared feature learning while losing information on sudden fluctuations of individual loads; single deep learning architectures (basic Transformer, shared decoder LSTM, etc.) lack the ability to adapt to the differences in characteristics of multiple loads and the guarantee of temporal causality, or they adopt a "one-size-fits-all" approach. The shared decoder of "cut" cannot adapt to the difference between high-frequency fluctuations in electrical load and large inertia in thermal load. It either "peeks" at future information due to the lack of causal mask in the attention mechanism, which violates the prediction timing logic. General optimization algorithms (such as the traditional Grey Wolf algorithm IGWO and grid search) are often inefficient in model hyperparameter tuning because they ignore architectural constraints. They either fail to ensure that the core constraint of "embedded dimension (embed_dim) is divisible by the number of attention heads (num_heads)" in Transformer-type models is generated as invalid parameters, or the computational cost increases due to full space traversal, making it difficult to efficiently find the optimal combination of hyperparameters that adapts to diverse loads.

[0025] The following provides a detailed description of the specific implementation methods of the embodiments of this application: Please see Figure 2 , Figure 2This is a flowchart illustrating the multi-element load forecasting method for integrated energy systems provided in this application embodiment. The multi-element load forecasting method for integrated energy systems can be implemented by server 110 and / or terminal 140. Figure 2 The integrated energy system multi-element load forecasting method shown includes: Step 210: Obtain the raw data of the integrated energy system, including electrical load, heat load, cooling load and meteorological data; Step 220: The electrical load, the heat load, the cooling load, and the meteorological data are fused to obtain a comprehensive dataset. The comprehensive dataset includes load sequence characteristics for four seasons. The load sequence characteristics for each season include electrical load sequence characteristics, heat load sequence characteristics, and cooling load sequence characteristics. Step 230: Determine multiple candidate features in the comprehensive dataset, wherein the candidate features are features associated with the electrical load, the thermal load, and the cooling load; Step 240: Among the candidate features, identify the N target features that are most relevant to the electrical load, the heat load, and the cooling load, where N is a positive integer greater than or equal to 1; Step 250: Use the target features as input to the preset load prediction model, and combine them with the output of the load prediction model to iteratively train the load prediction model using the JSSAO algorithm to obtain the optimal hyperparameters for the load prediction model output by the JSSAO algorithm. Step 260: Input the target features into the load prediction model configured with the optimal hyperparameters, and obtain enhanced features through forward dilation convolution and backward reverse dilation; Step 270: Process the enhanced features through the masked multi-head attention layer of the load prediction model to obtain the temporal feature vector; Step 280: Input the time-series feature vector into the load decoder of the load prediction model to obtain the multivariate load prediction results related to the electrical load, the heat load, and the cooling load output by the load decoder.

[0026] First, the overall architecture and implementation logic of this application are as follows: This application proposes a user-level integrated energy system multi-element load forecasting model based on STL-MIC-RBiTC-MMHAtt-JSSAO (e.g., ... Figure 3The model takes hourly load data and meteorological data as input and outputs three types of load forecasts for the next time step. The model first integrates the original data and divides the dataset according to season (spring March-May, summer June-August, autumn September-November, winter December-February). MinMaxScaler is used to standardize the data to the [0,1] interval. The dataset is then split into training and testing sets at an 8 / 2 ratio, and a time-series structure of "24 time-step input → 1 time-step output" is constructed, resulting in an input tensor of [number of samples, feature dimension, 24] and an output tensor of [number of samples, 3]. Feature engineering includes adding basic time features (holidays, hours, months, etc.) and time-related features (lag features, rolling features, periodic encoding). Then, STL decomposition is used to split each load into trend, seasonal, and residual components to form an initial feature set. The maximum information coefficient (MIC) is calculated using the minepy library, and the average MIC is used as the feature importance score. The Top-18 core features are selected to obtain a core feature tensor of [number of samples, 18, 24]. The core modeling layer is an RBiTC-MMHAtt collaborative architecture: The RBiTC layer receives the core feature tensor and outputs local enhancement features of [training batch size, embedding dimension, 24] through forward dilation convolution, backward inverse dilation convolution, and residual connections; the MMHAtt layer uses [24, training batch size, embedding dimension] after the feature dimension is transformed as Q / K, and the original core feature is transformed by 1×1 projection as [24, training batch size, embedding dimension] as V. Combined with the lower triangular causal mask, multi-dimensional dependencies are learned through 4 parallel attention heads, and the [training batch size, embedding dimension] feature of the last time step is taken as input to the decoder. Meanwhile, the JSSAO hyperparameter optimization module optimizes six core hyperparameters (embedding dimension 32-128, number of attention heads 2-8, convolution kernel size 3-7, number of training epochs 20-50, training batch size 32-128, learning rate 1e-4-1e-3): a valid initial population is generated using Latin hypercube sampling, and explorers (global search) and developers (local optimization) are dynamically split, outputting the optimal parameters after 10 iterations. Finally, multiple independent decoders with different loads are used: the electrical load decoder adapts to high-frequency fluctuations without Dropout, while the cold / hot load decoders add 0.01 Dropout to alleviate overfitting, and the output is a [training batch size, 3] prediction tensor; the predicted values ​​are obtained after training and inverse normalization using the Adam optimizer and MSE loss.

[0027] In step 210, data preprocessing and time-series sample construction (preprocessing of raw data) are performed. First, raw data from the user-level integrated energy system are collected, including hourly interval data on electrical load (kW), cooling load (CHWTON_kW, after unit conversion), and heating load (HTmmBTU_kW, after unit conversion). Meteorological data (temperature, air pressure, relative humidity, etc.) and time features (hours, months, holiday markers, etc.) are integrated simultaneously and merged into a comprehensive dataset by aligning with timestamps, removing invalid data with no temporal overlap. Subsequently, the dataset is divided into seasonal sub-datasets according to seasonal division rules (spring March-May, summer June-August, autumn September-November, winter December-February, with winter data integrated across years). MinMaxScaler is used to standardize the data to the [0,1] interval. The training and test sets are split in an 8:2 ratio, and an input tensor (containing features from the past 24 hours) and an output tensor (containing features from the past 24 hours) with dimensions of [number of samples, feature dimension, 24] and [number of samples, 3] are constructed according to the structure of "24 time-step input → 1 time-step output".

[0028] In step 220, STL-MIC fusion feature engineering is performed (generation of the comprehensive dataset). For each seasonal subset (sub-data corresponding to the four seasons), STL time-series decomposition is first performed on the electricity, cooling, and heating load sequences (with a period set to 24 to adapt to the daily load periodicity), obtaining the load sequence features for the four seasons. Each load sequence is split into a trend component (long-term variation pattern), a seasonal component (daily periodic fluctuation), and a residual component (random fluctuation), forming an initial feature set containing the original features and the decomposed components. Then, the maximum information coefficient (MIC) of each candidate feature with the electricity, cooling, and heating loads is calculated, and the average value is taken as the importance score. The Top-18 core features are selected in descending order of the scores, finally obtaining a core feature tensor (target feature) with dimensions [number of samples, 18, 24], filtering out low-association redundant information.

[0029] In step 230, multiple candidate features are determined in the comprehensive dataset, the candidate features being features associated with the electrical load, the thermal load, and the cooling load.

[0030] In step 240, N target features with the highest correlation to the electrical load, the thermal load, and the cooling load are determined from among the candidate features, where N is a positive integer greater than or equal to 1.

[0031] In some embodiments, determining the N target features that are most relevant to the electrical load, the heat load, and the cooling load from among the candidate features includes: For each candidate feature, a comprehensive correlation score is determined relating the candidate feature to the electrical load, the thermal load, and the cooling load. The top N candidate features with the highest comprehensive correlation scores among all the candidate features are selected as the target features.

[0032] Specifically, STL decomposition (splits the original data into load sequence features for the four seasons) STL decomposition is a time series decomposition method based on locally weighted regression (Loess) for data smoothing. Its core idea is to effectively decompose complex time series into three components: trend component, seasonal component, and residuals. It achieves better decomposition results, especially when dealing with seasonal data. The trend component represents the long-term trend of the time series; the seasonal component represents the periodic fluctuations of the time series; and the residuals represent the remaining fluctuations after removing the trend and periodicity (usually considered noise). STL decomposition can use an additive or multiplicative model. (1) (2) Seasonal fluctuations in energy load (electricity, cooling, and heating loads) are typically relatively stable in amplitude, with trends and seasonal fluctuations independent of each other. The fluctuation amplitude of the seasonal component St is unaffected by the trend component Tt. The additive model can more accurately separate stable periodic patterns. The trend Tt, seasonal St, and residual Rt components decomposed by the additive model are all absolute values ​​of the same order of magnitude as the original load: the trend component directly reflects the long-term trend of load changes, showing that the overall cooling load is higher in summer than in spring; the seasonal component directly reflects the load fluctuation within the daily cycle, directly showing how many kW the daytime load is higher than the nighttime load. This feature form is easier to integrate with other features (such as temperature, lagging load, time coding, etc.), reducing the learning difficulty of subsequent deep models (RBiTC + multi-head attention), making them more efficient at capturing the periodicity and trend patterns of the load. In this multi-energy load forecasting algorithm, the STL usage process is as follows: S1: Determine the decomposition objects and parameters. For the three types of multi-energy load time series, namely "electric load (KW)", "cooling load (CHWTON_kW)" and "heating load (HTmmBTU_kW)", the decomposition period is set to 24 to fit the daily periodicity of energy load. Each type of load series is used as the input for STL decomposition, and the trend, seasonality and residual components are extracted as the core input for subsequent feature engineering. S2: Seasonal STL Feature Extraction. Due to the significant seasonal differences in energy load characteristics, the merged load-meteorological data is first split into four seasons: spring, summer, autumn, and winter. STL decomposition is then performed on the three types of load for each season. For example, for the summer electricity load sequence, STL decomposition yields KW_trend (long-term electricity consumption trend), KW_seasonal (daily periodic electricity consumption fluctuation), and KW_resid (residual noise). Similarly, decomposition components for cooling and heating loads are generated, ultimately forming a feature set named "Load_Trend / Season / Residual". S3: Multi-source feature fusion. The three types of component features obtained from STL decomposition are fused with meteorological features, load lag features (such as KW_lag1-KW_lag3), rolling statistical features (such as KW_roll24 mean and KW_roll168 variance), and periodic coding features (sine / cosine coding for hours and months) to construct a feature matrix with rich dimensions, providing a foundation for subsequent feature selection and model training; Here, the maximum information coefficient (MIC) feature is used for screening (the maximum information coefficient is also the comprehensive association score described in this application). The Maximum Information Coefficient (MIC) is a statistic based on mutual information theory. Through grid partitioning and probability distribution analysis, it measures the strength of the linear or nonlinear correlation between two variables; the larger the value and the closer it is to 1, the stronger the correlation. In this multi-energy load forecasting algorithm, MIC is used to filter features highly correlated with electricity, cooling, and heating loads, providing high-quality input for the deep model. Its implementation in the algorithm is as follows: S1: Determine the target and calculation object for feature selection. Clearly define the input features to be screened that are highly correlated with the three types of multi-energy loads: electrical load, cooling load, and heating load. Candidate features include STL decomposition features (trend, seasonality, residual), meteorological features (temperature, wind speed, etc.), lag features, rolling statistical features, periodic coding features, etc., forming a feature set to be evaluated; S2: MIC Calculation and Multi-Objective Fusion. For each candidate feature, its MIC value with a single load objective is calculated using the MINE class of the minepy library. The MIC calculation is based on the grid partitioning principle of mutual information, and the formula is: (3) Where a and b are the number of grid divisions along the x and y axes, S is the upper limit of the total number of grids (in practical applications, it is taken as 0.6 times the number of data samples), and I(f(x), f(y)) is the mutual information between the feature and the target. For each feature, the MIC value with the electrical, cooling, and heating loads is calculated and averaged to serve as the comprehensive association score for that feature, thus quantifying the association of multiple target features. S3: Feature Selection and Model Input Construction. All candidate features are sorted from high to low according to their MIC comprehensive scores, and the Top 18 features (N is 18 in this application) are selected as the input for the subsequent deep prediction model. These features contain strong relevant information directly related to the load, and avoid the limitations of single load correlation through multi-objective fusion, providing high-quality input for the joint prediction of multi-energy loads.

[0033] In step 250, the target feature is used as the input of a preset load prediction model, and combined with the output of the load prediction model, the load prediction model is iteratively trained using the JSSAO algorithm to obtain the optimal hyperparameters for the load prediction model output by the JSSAO algorithm.

[0034] Here, the core feature modeling of RBiTC-MMHAtt is performed. The core feature tensor is first input into the RBiTC (Residual Bidirectional Temporal Convolution) layer: the forward dilation convolution captures the local dependencies of "past → present", and the backward reverse dilation convolution (input reverse order → convolution → output reverse order) captures the reverse local correlations of "present → past". The two are added to the residual features after 1×1 convolution, and the output is a local enhancement feature of [number of samples, embed_dim, 24], which alleviates gradient vanishing and preserves local details (such as load peaks). Enhanced features are then input into the MMHatt (masked multi-head attention) layer: The RBiTC output is transformed into [24, number of samples, embed_dim] as Q (query) and K (key) through dimensional transformation. The original core features are transformed into [24, number of samples, embed_dim] through 1×1 projection as V (value, preserving burst features). Combined with the lower triangular causal mask and learning multi-dimensional global dependencies through 4 parallel attention heads, the features at the last time step are taken to obtain the temporal feature vector of [number of samples, embed_dim].

[0035] In some embodiments, an initial population is obtained by initializing the population of hyperparameters for the load forecasting model through Latin hypercube sampling. The initial population includes multiple individuals, each representing a set of hyperparameter combinations. The model prediction value for each individual is determined by the load prediction model, and the fitness score for each individual is determined based on the model prediction value for each individual. The initial population is divided into an exploratory subgroup and a development subgroup. The exploratory subgroup focuses on global search, while the development subgroup focuses on local fine-grained search. The exploration subgroup and the development subgroup are iteratively trained, and the population division ratio of the exploration subgroup and the development subgroup is dynamically adjusted with the number of iterations; After a preset number of iterations of training, the individual with the highest fitness score in the merged subgroup formed by the exploration subgroup and the development subgroup is selected as the optimal individual. The combination of hyperparameters corresponding to the optimal individual is taken as the optimal hyperparameters.

[0036] Specifically, the snow ablation optimization algorithm improved by combining the oppositional selection operator and random perturbation (JSSAO algorithm) is proposed. SAO (Snow Sublimation and Melting) is a new generation of global optimization algorithm based on the sublimation and melting behavior of snow, possessing powerful optimization capabilities and fast convergence speed. However, the original SAO algorithm forms an unevenly distributed population during initial random initialization, leading to a decrease in population diversity. As the number of iterations increases, the population quality and the algorithm's global exploration and local exploration capabilities continuously weaken, making the algorithm prone to getting trapped in local optima. This algorithm enhances the original SAO's ability to balance global search and local exploitation by integrating the Joint Opposite Selection (JOS) operator and a random perturbation mechanism. It also optimizes the population partitioning strategy, dynamically dividing the population into an "exploratory subgroup" simulating the snow sublimation process and focusing on global search, and an "exploratory subgroup" simulating the snow melting process and focusing on local fine-grained search. The subgroup ratio is adjusted as the iteration progresses—the size of the exploration subgroup gradually decreases, and the size of the exploration subgroup gradually increases, thereby balancing the efficiency of global exploration and local exploitation and accelerating algorithm convergence. The algorithm first initializes the population using Latin hypercube sampling. For each parameter dimension (denoted as D), its value range is divided into N equal-width intervals. One sample point is randomly generated in each interval, and these intervals are combined to form the initial population. (4) Each individual This represents a set of hyperparameter combinations. The fitness function is the coefficient of determination of the deep model on the validation set. As evaluation indicators, i.e. (5) in (Xi) represents the model prediction value corresponding to the hyperparameter Xi. A higher fitness score (as described in this application) indicates a better combination of hyperparameters. During the iteration process, the population partitioning ratio is dynamically adjusted with the number of iterations. Let the maximum number of iterations be... The threshold for separating the explorers and developers subgroups at the t-th iteration is: (6) That is, the size of the exploratory subgroup gradually shrinks with each iteration, while the size of the development subgroup expands accordingly. For the exploratory subgroup, its position update integrates the Joint Opposites Selection (JOS) operator, Brownian motion, and random perturbation. The Joint Opposites Selection (JOS) operator is introduced, generating oppositional solutions through selective leadership opposition and dynamic opposition, effectively expanding the population's coverage in the solution space, avoiding local clustering problems caused by random initialization of the initial population, and significantly improving global exploration capabilities. A random perturbation mechanism is added, introducing normally distributed random noise during individual position updates, enhancing the randomness of local search in later iterations, and helping the algorithm escape local optima traps. The position update formula for the exploratory subgroup is: (7) JOS The complementary solution generated by the JOS operator is BM(t). For Brownian motion random quantities, This is the current globally optimal solution. The population mean These are the weighting coefficients. For random perturbations ( As the iteration decreases, the global search range is expanded; for the developing subgroup, its position update focuses more on approaching the global optimum, as shown in the formula. (8) in The optimization accuracy is improved through local fine-grained search. After each iteration, the updated subgroups are merged, and the top N individuals are retained according to their fitness to enter the next iteration, until the maximum number of iterations is reached. Finally, the hyperparameter combination corresponding to the individual with the highest fitness is output as the optimal solution. The implementation steps of JSSAO in the model are as follows: S1: Parameter space definition. Determine the hyperparameters and range of the deep model to be optimized, including embedding dimension (embed_dim), number of attention heads (num_heads), kernel size (kernel_size), number of training epochs (epochs), batch size (batch_size), and learning rate (lr), forming the parameter space (param_space); S2: Population Initialization. The `latin_hypercube_sampling` function is called to generate an initial population (size 6) through Latin hypercube sampling. This ensures that individuals are uniformly distributed in the parameter space. Simultaneously, the validity of the parameters is corrected after initialization, guaranteeing that the embedding dimension (`embed_dim`) must be an integer multiple of the number of heads in the multi-head attention mechanism (`num_heads`). The multi-head attention mechanism requires the embedding dimension to be evenly distributed across each "head," with each head independently processing a portion of the features. If the embedding dimension is not divisible by the number of heads, it cannot be evenly distributed, leading to a dimension mismatch. S3: Initial fitness calculation. For each initial individual (hyperparameter combination), call the function to train the deep model and calculate the model's fitness on the validation set. As a fitness metric, the initial optimal individual is selected.

[0037] S4: Iterative optimization. The population is dynamically divided into explorers and developers based on the number of iterations. The position update of the explorers combines the JOS operator to generate complementary solutions, and the search range is expanded through Brownian motion and random perturbation. The position update of the developers moves towards the global optimum, and the search is refined through local perturbation. The fitness of the updated individuals is calculated, the subgroups are merged and sorted by fitness, and the best N individuals are retained.

[0038] S5: Termination and Output. After reaching the maximum number of iterations, the hyperparameter combination corresponding to the individual with the highest fitness is output as the optimal hyperparameters for the deep model. The flowchart of this algorithm is as follows. Figure 4 As shown.

[0039] In step 260, the target feature is input into the load prediction model configured with the optimal hyperparameters, and enhanced features are obtained through forward dilation convolution and backward dilation. Specifically, forward dilation convolution is performed on the input sequence corresponding to the target feature to obtain the forward dilation convolution result, so as to capture the local time-dependent learning load continuous change trend. The input sequence corresponding to the target feature is subjected to backward reverse expansion to obtain the backward expansion convolution result, so as to capture the recovery trend after the peak of reverse local dependency learning; The enhanced feature is obtained by adding the target feature to the forward dilation convolution result and the backward dilation convolution result.

[0040] In step 270, the enhanced features are processed by the masked multi-head attention layer of the load prediction model to obtain a temporal feature vector.

[0041] Specifically, Masked Multi-Head Attention (MMHAtt). MMHAtt's core function is to accurately capture the local temporal dependencies of multiple energy loads in time series forecasting scenarios. At the same time, it uses causal masking to ensure that the forecasting process does not leak future information, laying the foundation for generating reliable forecasting results by combining the encoder's global context.

[0042] For the sequence data input to the decoder, feature mapping and multi-head splitting are required first. Let the input sequence of a certain layer of the decoder be... (where l is the sequence length and d is the feature dimension), through three sets of learnable linear projection matrices ( (For the dimensions of query and key), map the input sequence to the query matrix Q, key matrix K, and value matrix V respectively: (9) To capture the dependencies between different feature subspaces, MMHatt further decomposes (Q, K, V) into h independent subspaces (i.e., attention heads), each subspace corresponding to a dedicated set of projection matrices. ( Within each attention head, attention weights are assigned based on the similarity between the query and the key, and the values ​​are weighted and summed to obtain the attention output for that head. To avoid... An excessively large scaling factor will result in excessively high variance in the dot product; therefore, a scaling factor needs to be introduced. The formula for calculating single-head attention is: (10) Because multi-energy load forecasting needs to adhere to temporal causality (i.e., current forecasts rely solely on historical data), MMHatt introduces a lower triangular causal mask Mm before Softmax calculation (the elements corresponding to future time steps in the mask matrix are...). After Softmax, the weights approach 0. At this point, the single-head attention formula is corrected to: (11) Finally, the outputs of the h attention heads are combined through a concatenation operation and then processed by a set of learnable projection matrices. Mapping back to the original feature dimensions yields the final output of MMHatt. (12) By enriching the dimensions of dependency capture through a multi-head parallel mechanism and strictly adhering to the logical constraints of time-series prediction with the help of causal masks, MMHatt can effectively adapt to the modeling requirements of local time-series characteristics of multi-energy loads in user-side integrated energy systems.

[0043] Masked Multi-Head Attention Mechanism Based on Inverse Bidirectional Temporal Convolution (RBiTC-MMHAttt) Traditional multi-energy load forecasting models primarily capture coupling relationships within a global environment, but tend to overlook the unique local load behavior of each task. To address this issue, the proposed load forecasting model introduces the RBiTC-MMHAttt mechanism to improve sensitivity to local load characteristics. The MMHAttt mechanism in the decoder, utilizing a dot-product self-attention mechanism, is crucial for effectively capturing the temporal dependencies of each load. However, due to its uniform attention distribution and limited context window, this mechanism has limited sensitivity to local environments, potentially leading to inaccurate peak predictions and difficulty in capturing sudden spikes common in integrated energy systems on the user side. Figure 5 The similarity between Q and K in the traditional MMHattt mechanism shown is calculated using a single point value, which fails to effectively utilize the local sequence context. It does not consider that QK matching of local time series may mask the change points in the MMHattt mechanism, thus limiting the accuracy of multi-energy load forecasting.

[0044] To overcome this problem, this application adds an RBiTC layer in MMHattt to handle the input of the task-specific decoder, such as... Figure 6 As shown in the figure, (Conv,1) represents a causal convolution with a kernel size of 1, while (RBiTC,K) represents an RBiTC with a kernel size of K. This enhancement significantly improves the integration of local contextual information in the self-attention mechanism. By using the RBiTC layer, each decoder can expand its receptive field, thereby integrating richer contextual information without exponentially increasing complexity. This allows the proposed multi-energy load prediction model to maintain fine sensitivity to local context while adhering to the temporal causality necessary for multi-energy load prediction tasks. This improvement not only accurately predicts the overall pattern but also the peak values ​​in each multi-energy load. Causal convolution applies time-series convolution in the temporal domain by using a one-dimensional dilated causal convolution kernel. The output of each layer comes only from the information of the previous layer and the corresponding temporal inputs of the current and past time steps, ensuring strict adherence to temporal constraints. In a given time series... This represents the time series of the current location, while Indicates in The previously input time series is defined as follows. The convolution kernel is defined as F=[f1,f2,…,fk], and the time series is defined as X= The dilated causal convolution operation can then be expressed as: (13) Where k represents the size of the convolution kernel, sd i represents the past direction, and d is the dilation factor used to specify the spacing between adjacent convolutional kernel elements determined by inserting zero vectors. Dilated causal convolution, limited to forward convolution on the input sequence, can only extract forward load data, ignoring valuable backward information. Therefore, this prediction model introduces bidirectional causal convolution to capture temporal features from both the forward and backward directions, thereby enhancing the ability of each task-specific decoder to capture long-term dependencies in the load sequence and more effectively acquire local information. Backward information enables the model to identify temporal dependencies and patterns presented in reverse order, broadening the prediction context by providing an understanding of the impact of future events on past and current states. Understanding the impact of factors such as surges in energy demand, long holidays, or extreme weather events can significantly improve the accuracy of load forecasting. Furthermore, backward information aids in error correction and anomaly detection, leading to more detailed and accurate modeling of the local environment.

[0045] Because the bidirectional causal convolution employed achieves multiple-pass convolution, the model receives a larger receptive field. However, expanding the receptive field introduces challenges such as gradient vanishing and slow convergence. To address this new problem brought about by introducing bidirectional causal convolution, residual blocks are introduced into the model to develop RBiTC, thereby enabling efficient feature extraction from sequences. The residual elements can be represented as: (14) in This represents the final output of the residual block. Before each decoder executes the MMHatt mechanism, the input sequence is processed using a convolutional kernel of size K to generate Q and K in the RBiTC layer. The causal nature of this convolution ensures that the current time step is not affected by future information. The Q and K generated by RBiTC have a higher understanding of the surrounding local time series, thus enabling similarity calculation based on local context information and improving prediction accuracy. When the convolutional kernel size K is set to 1, the convolutional self-attention mechanism simplifies to the traditional self-attention mechanism, demonstrating its ability to generalize based on kernel size changes. The R-MMHAtt mechanism proposed in this study combines the advantages of RBiTC with self-attention, aiming to effectively capture strong temporal dependencies in time series data while considering potential anomalies. To solve the multi-energy load prediction problem, the input sequence is first processed by RBiTC to extract the contextual information around each time step, and then a self-attention mechanism is applied to model the dependencies between different time points. In this process, both the Q and K matrices are generated from the convolutional output, while the V matrix retains the inherent characteristics of the original sequence. By adjusting the bidirectional convolutional kernel size K, the proposed R-MMHAtt mechanism can dynamically optimize its focus on local temporal patterns at different ranges. Unlike traditional Transformer networks that primarily emphasize long-range dependencies, this method preserves temporal causality while maintaining sensitivity to changes in local context. The specific steps of the RBiTC-MMHAtt mechanism implementation in the model are as follows: Step 1: Preparation of Input Features and Hyperparameters S1: Feature Input. The Top 18 key features (including STL decomposition features, lag load, meteorological factors, etc.) obtained from the MIC feature screening in Section 3.2 are used as the input feature set of the mechanism to ensure that the input contains information strongly correlated with electricity, cooling, and heating loads.

[0046] S2: Hyperparameter Configuration. Load the optimal hyperparameters obtained from the JSSAO optimization in Section 3.3, including the embedding dimension (embed_dim), the number of multi-head attention heads (num_heads), and the kernel size (kernel_size). Embed_dim must be an integer multiple of num_heads (to ensure the validity of multi-head feature allocation), and kernel_size must be an odd number (to adapt to the bidirectional convolution of RBiTC).

[0047] S3: Data format conversion. The input features are reshaped into a 3D tensor [number of samples, number of features, sequence length] (the sequence length is fixed at 24, corresponding to a 24-hour time series window), where the "number of features" is 18, to adapt to the convolutional input format of the RBiTC layer and the sequence modeling requirements of MMHatt.

[0048] Step 2: Constructing the RBiTC layer to enhance the learning of local context. S1: Layer structure configuration. Initialize the RBiTC layer with the following parameters: number of input channels = 18 (Top 18 features), number of output channels = embed_dim (JSSAO optimized value), kernel size = kernel_size (JSSAO optimized value), and dilation = 2.

[0049] S2: Bidirectional convolution calculation.

[0050] Forward dilation convolution: Dilate the input sequence using convolution (conv_forward1→ReLU→conv_forward2→ReLU) to capture the continuous change trend of local time-dependent learning load from "past to present"; Inverse dilated convolution: First, reverse the input sequence, then perform the same operation as the forward convolution, and finally reverse the result to capture the recovery trend after the peak of inverse local dependency learning from "current to future". S3: Residual Connection. The original input features are mapped to the embed_dim dimension via a 1×1 convolution, and then added to the results of the forward + backward convolutions. This alleviates the gradient vanishing problem in deep networks and enhances local feature flow.

[0051] Step 3: MMHatt layer construction, capturing global dependencies. S1: Multi-head attention configuration. Initialize the masked multi-head self-attention layer with the following parameters: embedding dimension = embed_dim, number of attention heads = num_heads (JSSAO optimized value). Learn the feature dependencies of different subspaces through a multi-head parallel mechanism.

[0052] S2: Q, K, V are generated.

[0053] Q (query) and K (key): Taken from the output of the RBiTC layer, and transformed into [seq_len, batch_size, embed_dim] to ensure that attention similarity is calculated based on local context; V (value): Taken from the result of the original input features after 1×1 convolution projection, the dimension is converted to [seq_len, batch_size, embed_dim], and the inherent characteristics of the original features are preserved; S3: Application of causal masking. Generates a lower triangular mask that only allows the current time step to focus on the past and the present moment, preventing the leakage of future information.

[0054] Step 4: Feature Fusion and Multi-Load Output S1: Core Feature Extraction. The features from the last time step output of the MMHatt layer are taken as the core features after fusing global dependencies and local context.

[0055] S2: Task-specific decoding. The core features are mapped to predicted values ​​of electrical load (KW), cooling load (CHWTON_kW), and heating load (HTmmBTU_kW) through three independent fully connected layers (fc_kw, fc_chw, fc_ht) respectively (each output dimension is 1), and concatenated to form the final prediction result of [batch_size, 3], realizing the joint prediction of multi-energy load.

[0056] This mechanism expands the local receptive field and enhances the sensitivity to peaks / spiks through the RBiTC layer, and captures global time dependence by combining the MMHatt layer. The two work together to achieve local fine modeling and global correlation capture. Finally, based on the high-quality features selected by MIC and the hyperparameters optimized by JSSAO, the accuracy of multi-energy load prediction is improved.

[0057] In step 280, the time-series feature vector is input into the load decoder of the load prediction model to obtain the multivariate load prediction results related to the electrical load, the heat load, and the cooling load output by the load decoder.

[0058] Specifically, the time-series feature vector is input into the electrical load decoder, the thermal load decoder, and the cold load decoder respectively to obtain the electrical load prediction result output by the electrical load decoder, the thermal load prediction result output by the thermal load decoder, and the cold load prediction result output by the cold load decoder. The multi-load forecast result is determined based on the electrical load forecast result, the heat load forecast result, and the cooling load forecast result.

[0059] Model performance evaluation To accurately analyze the prediction performance of the proposed model on multi-energy loads (electrical load, cooling load, and heating load), this paper selects root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²) as evaluation indicators. The first three are used to measure the magnitude of the prediction error (the smaller the value, the better), while R² is used to measure the goodness of fit of the model (the closer the value is to 1, the better). The calculation method is as follows: (15) (16) (17) (18) in This represents the actual load value. To predict load values, is the mean of the true values, and n is the sample size.

[0060] The innovative aspects of this application are as follows: 1. Design of Reverse Bidirectional Temporal Convolution (RBiTC). This application solves the problem that traditional causal convolution can only capture unidirectional temporal dependencies and that gradient vanishing is caused by expanding the receptive field by bidirectionally fusing forward and backward dilated convolutions and combining residual connections. This enhances the model's sensitivity to local load characteristics and enables fine modeling of local context.

[0061] 2. Construction of Multi-Head Hybrid Attention (MMHAtt) Mechanism. This application generates Q and K (integrating local context) based on RBiTC output, and generates V (preserving inherent characteristics) from the original features. By combining multi-head parallel attention and causal masking, it solves the problem of traditional self-attention relying on single-point QK matching and ignoring local sequence context. It balances the association learning of global time dependence and local features, and improves the ability to capture the coupling relationship of multiple energy loads.

[0062] 3. Synergy between MIC feature selection and JSSAO hyperparameter optimization. This application uses MIC to calculate the comprehensive correlation score between features and multi-energy loads to select top key features and solve the feature redundancy problem; it uses the JSSAO algorithm to dynamically divide the exploration / development subgroups to optimize the model hyperparameters, solving the problems of inefficient hyperparameter tuning and easy getting trapped in local optima, thereby improving the quality of input features and the rationality of model configuration.

[0063] 4. Seasonally Adaptive Single-Load Modeling Strategy. To address the significant differences in electricity, cooling, and heating load characteristics across the four seasons, the multi-task shared decoder is replaced with a single-load independent decoder. This resolves the insufficient adaptability of the shared model in scenarios with differentiated characteristics, improving the stability and accuracy of all-season forecasts.

[0064] In summary, existing multi-energy load forecasting technologies generally suffer from problems such as incomplete capture of local context, imbalance between global dependence and local characteristic modeling, low efficiency in feature selection and hyperparameter optimization, and insufficient adaptability to scenarios with large differences in load characteristics. These technologies mostly rely on unidirectional time-series modeling to capture local features, have limited utilization of local sequence context by attention mechanisms, and lack adaptive designs for the differentiation of multi-energy load characteristics, resulting in limited prediction accuracy and stability in complex scenarios such as sudden spikes and irregular fluctuations in load patterns. This application systematically addresses these shortcomings through multi-dimensional technological innovation, achieving a comprehensive improvement in prediction performance.

[0065] (1) Bidirectional local feature enhancement and global dependency collaborative modeling. To address the problem of unidirectional capture of local dependencies and imbalance between global and local modeling in existing technologies, this application introduces the RBiTC-MMHAtt collaborative mechanism: The RBiTC layer captures bidirectional local temporal dependencies through bidirectional dilated convolution and residual connections, expands the receptive field and avoids gradient vanishing, and enhances the sensitivity to local load characteristics; the MMHAtt layer uses the RBiTC output as Q / K and the original feature projection as V, and combines causal masking to enable attention to be calculated based on local sequence context, thereby achieving accurate modeling of global dependencies.

[0066] (2) Highly adaptable features and hyperparameter co-optimization. To address the problems of feature redundancy and inefficient hyperparameter tuning in existing technologies, this application adopts a co-optimization scheme of MIC feature selection and JSSAO hyperparameter optimization. Top key features (target features) are selected by MIC (maximum information coefficient) to reduce interference from irrelevant information. The JSSAO algorithm integrates the generation of contradictory solutions and random perturbation mechanism to optimize the model hyperparameters, ensuring that the hyperparameter configuration is adapted to the complex patterns of multi-energy loads and improving the model training efficiency and generalization ability.

[0067] (3) Seasonal / Load Characteristic Differentiation Adaptation Design. To address the problem that existing multi-task shared models are difficult to adapt to large differences in load characteristics, this application designs a single-load independent modeling strategy. The original multi-task shared decoder is replaced with a single-load dedicated decoder to build independent prediction branches for electricity, cooling and heating loads respectively. Each branch focuses on learning the unique patterns of the corresponding load to avoid feature confusion caused by conflicts in different load characteristics, thereby improving the prediction adaptability of seasonal transition periods or characteristic differentiation scenarios.

[0068] Please see Figure 7 , Figure 7 This is a schematic diagram of the structure of the integrated energy system multi-element load forecasting device provided in the embodiments of this application. The integrated energy system multi-element load forecasting device is applied to computer equipment, and the integrated energy system multi-element load forecasting device may include: Acquisition unit 401 is used to acquire raw data of the integrated energy system, including electrical load, heat load, cooling load and meteorological data; The fusion unit 402 is used to fuse the electrical load, the heat load, the cooling load and the meteorological data to obtain a comprehensive dataset. The comprehensive dataset includes load sequence characteristics for four seasons, and the load sequence characteristics for each season include electrical load sequence characteristics, heat load sequence characteristics and cooling load sequence characteristics. The first determining unit 403 is used to determine multiple candidate features in the comprehensive dataset, wherein the candidate features are features associated with the electrical load, the thermal load and the cooling load; The second determining unit 404 is used to determine, among the candidate features, the N target features that are most relevant to the electrical load, the thermal load and the cooling load, where N is a positive integer greater than or equal to 1; Training unit 405 is used to take the target features as input to a preset load prediction model, and combine them with the output of the load prediction model to iteratively train the load prediction model using the JSSAO algorithm to obtain the optimal hyperparameters for the load prediction model output by the JSSAO algorithm. Input unit 406 is used to input the target features into the load prediction model configured with the optimal hyperparameters, and obtain enhanced features through forward dilation convolution and backward reverse dilation; Enhancement unit 407 is used to process the enhanced features through the masked multi-head attention layer of the load prediction model to obtain a temporal feature vector; The prediction unit 408 is used to input the time-series feature vector into the load decoder of the load prediction model to obtain the multi-variable load prediction results related to the electrical load, the heat load, and the cooling load output by the load decoder.

[0069] Reference Figure 8 , Figure 8 To implement the structural block diagram of a portion of the terminal 140 in this application embodiment, the terminal 140 includes: a radio frequency (RF) circuit 710, a memory 715, an input unit 730, a display unit 740, a sensor 750, an audio circuit 760, a wireless fidelity (WiFi) module 770, a processor 780, and a power supply 790, among other components. Those skilled in the art will understand that... Figure 8 The terminal 140 structure shown does not constitute a limitation on a mobile phone or computer, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0070] The RF circuit 710 can be used to receive and transmit signals during information transmission or calls. In particular, it receives downlink information from the base station and processes it with the processor 780; in addition, it transmits uplink data to the base station.

[0071] The memory 715 can be used to store software programs and modules. The processor 780 executes various functional applications of the terminal and multi-load forecasting processing of the integrated energy system by running the software programs and modules stored in the memory 715.

[0072] The input unit 730 can be used to receive input numeric or character information, and to generate key signal inputs related to the terminal's settings and function control. Specifically, the input unit 730 may include a touch panel 731 and other input devices 732.

[0073] The display unit 740 can be used to display input or provided information, as well as various menus of the terminal. The display unit 740 may include a display panel 741.

[0074] Audio circuitry 760, speaker 761, and microphone 762 provide an audio interface.

[0075] In this embodiment, the processor 780 included in the terminal 140 can execute the integrated energy system multi-load forecasting method of the previous embodiment.

[0076] The terminal 140 in this application embodiment includes, but is not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, and aircraft. This application embodiment can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving.

[0077] Figure 9 This is a partial structural block diagram of a server 110 implementing an embodiment of this application. The server 110 can vary significantly due to different configurations or performance characteristics, and may include one or more central processing units (CPUs) 822 (e.g., one or more processors) and memory 832, and one or more storage media 830 (e.g., one or more mass storage devices) for storing application programs 842 or data 844. The memory 832 and storage media 830 can be temporary or persistent storage. The program stored in the storage media 830 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the server 110. Furthermore, the CPU 822 may be configured to communicate with the storage media 830 and execute the series of instruction operations in the storage media 830 on the server 110.

[0078] Server 110 may also include one or more power supplies 826, one or more wired or wireless network interfaces 850, one or more input / output interfaces 858, and / or one or more operating systems 841, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0079] The central processing unit 822 in server 110 can be used to execute the integrated energy system multi-load forecasting method of this application embodiment.

[0080] This application also provides a computer-readable storage medium for storing program code for executing the integrated energy system multi-load forecasting method of the foregoing embodiments.

[0081] This application also provides a computer program product, which includes a computer program. A processor of a computer device reads and executes the computer program, causing the computer device to perform the aforementioned integrated energy system multi-load forecasting method.

[0082] Furthermore, the terms “comprising” and “including”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0083] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0084] It should be understood that in the description of the embodiments of this application, "multiple" means two or more, "greater than", "less than", "exceeding" etc. are understood to exclude the number itself, and "above", "below", "within" etc. are understood to include the number itself.

[0085] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0086] The units described as separate components may or may not be physically separate. 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 units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0087] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0088] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0089] It should also be understood that the various implementation methods provided in this application can be combined arbitrarily to achieve different technical effects.

[0090] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0091] The above is a detailed description of the embodiments of this application. However, this application is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A multi-element load forecasting method for integrated energy systems, characterized in that, The method includes: Acquire raw data from the integrated energy system, including electrical load, heat load, cooling load, and meteorological data; The electrical load, heat load, cooling load, and meteorological data are fused to obtain a comprehensive dataset. The comprehensive dataset includes load sequence characteristics for four seasons, and the load sequence characteristics for each season include electrical load sequence characteristics, heat load sequence characteristics, and cooling load sequence characteristics. Multiple candidate features are identified in the comprehensive dataset, and these candidate features are those associated with the electrical load, the thermal load, and the cooling load. Among the candidate features, N target features with the highest correlation to the electrical load, the thermal load, and the cooling load are identified, where N is a positive integer greater than or equal to 1; The target features are used as input to a preset load forecasting model, and combined with the output of the load forecasting model, the JSSAO algorithm is used to iteratively train the load forecasting model to obtain the optimal hyperparameters for the load forecasting model output by the JSSAO algorithm. The target features are input into the load prediction model configured with the optimal hyperparameters, and enhanced features are obtained through forward dilation convolution and backward reverse dilation. The enhanced features are processed by the masked multi-head attention layer of the load prediction model to obtain a temporal feature vector; The time-series feature vector is input into the load decoder of the load prediction model to obtain the multivariate load prediction results related to the electrical load, the heat load, and the cooling load output by the load decoder.

2. The multi-source load forecasting method for integrated energy systems according to claim 1, characterized in that, The meteorological data includes temperature data, air pressure data, and relative humidity data. The electrical load, heat load, cooling load, and meteorological data are fused to obtain a comprehensive dataset, including: The electrical load, thermal load, cooling load, temperature data, air pressure data, and relative humidity data are fused to obtain fused data. The fused data is divided into four seasons to obtain sub-data corresponding to the four seasons; For each season, the sub-data corresponding to the season is decomposed into time series to obtain the electrical load sequence characteristics, heat load sequence characteristics, and cold load sequence characteristics corresponding to the season. The electrical load sequence characteristics, heat load sequence characteristics, and cold load sequence characteristics corresponding to the season are then fused with the temperature data, air pressure data, and relative humidity data corresponding to the season to obtain the fused characteristics corresponding to the season. The fusion features corresponding to each season are fused to obtain the comprehensive dataset.

3. The integrated energy system multi-source load forecasting method according to claim 1, characterized in that, The step of identifying N target features among the candidate features that are most relevant to the electrical load, the heat load, and the cooling load includes: For each candidate feature, a comprehensive correlation score is determined relating the candidate feature to the electrical load, the thermal load, and the cooling load. The top N candidate features with the highest comprehensive correlation scores among all the candidate features are selected as the target features.

4. The integrated energy system multi-source load forecasting method according to claim 1, characterized in that, The step involves using the target features as input to a pre-defined load forecasting model, and combining the output of the load forecasting model with the JSSAO algorithm to iteratively train the load forecasting model, thereby obtaining the optimal hyperparameters for the load forecasting model output by the JSSAO algorithm. This includes: The population of hyperparameters for the load prediction model is initialized by Latin hypercube sampling to obtain an initial population, which includes multiple individuals, each representing a set of hyperparameter combinations. The model prediction value for each individual is determined by the load prediction model, and the fitness score for each individual is determined based on the model prediction value for each individual. The initial population is divided into an exploratory subgroup and a development subgroup. The exploratory subgroup focuses on global search, while the development subgroup focuses on local fine-grained search. The exploration subgroup and the development subgroup are iteratively trained, and the population division ratio of the exploration subgroup and the development subgroup is dynamically adjusted with the number of iterations; After a preset number of iterations of training, the individual with the highest fitness score in the merged subgroup formed by the exploration subgroup and the development subgroup is selected as the optimal individual. The combination of hyperparameters corresponding to the optimal individual is taken as the optimal hyperparameters.

5. The multi-source load forecasting method for integrated energy systems according to claim 1, characterized in that, The step of inputting the target features into the load prediction model configured with the optimal hyperparameters, and obtaining enhanced features through forward dilation convolution and backward reverse dilation, includes: The input sequence corresponding to the target feature is subjected to forward dilation convolution to obtain the forward dilation convolution result, so as to capture the continuous change trend of local time-dependent learning load; The input sequence corresponding to the target feature is subjected to backward reverse expansion to obtain the backward expansion convolution result, so as to capture the recovery trend after the peak of reverse local dependency learning; The enhanced feature is obtained by adding the target feature to the forward dilation convolution result and the backward dilation convolution result.

6. The multi-source load forecasting method for integrated energy systems according to claim 1, characterized in that, The enhancement features are processed through the masked multi-head attention layer of the load prediction model to obtain a temporal feature vector, including: The input sequences corresponding to the enhanced features are respectively mapped to a query matrix Q, a key matrix K, and a value matrix V; The enhanced features are processed by the query matrix Q, the key matrix K, and the value matrix V to obtain the temporal feature vector.

7. The integrated energy system multi-source load forecasting method according to claim 1, characterized in that, The load decoder includes an electrical load decoder, a thermal load decoder, and a cold load decoder. The step of inputting the time-series feature vector into the load decoder of the load prediction model to obtain the multivariate load prediction results output by the load decoder related to the electrical load, the thermal load, and the cold load includes: The time-series feature vectors are respectively input into the electrical load decoder, the thermal load decoder, and the cold load decoder to obtain the electrical load prediction result output by the electrical load decoder, the thermal load prediction result output by the thermal load decoder, and the cold load prediction result output by the cold load decoder. The multi-load forecast result is determined based on the electrical load forecast result, the heat load forecast result, and the cooling load forecast result.

8. A multi-element load forecasting device for an integrated energy system, characterized in that, The device includes: The acquisition unit is used to acquire raw data of the integrated energy system, including electrical load, heat load, cooling load and meteorological data; The fusion unit is used to fuse the electrical load, the heat load, the cooling load, and the meteorological data to obtain a comprehensive dataset. The comprehensive dataset includes load sequence characteristics for four seasons, and the load sequence characteristics for each season include electrical load sequence characteristics, heat load sequence characteristics, and cooling load sequence characteristics. The first determining unit is configured to determine multiple candidate features in the comprehensive dataset, wherein the candidate features are features associated with the electrical load, the thermal load, and the cooling load; The second determining unit is used to determine, among the candidate features, the N target features that are most relevant to the electrical load, the thermal load and the cooling load, where N is a positive integer greater than or equal to 1; The training unit is used to take the target features as input to a preset load prediction model, and combine them with the output of the load prediction model to iteratively train the load prediction model using the JSSAO algorithm, so as to obtain the optimal hyperparameters for the load prediction model output by the JSSAO algorithm. The input unit is used to input the target features into the load prediction model configured with the optimal hyperparameters, and obtain enhanced features through forward dilation convolution and backward reverse dilation; An enhancement unit is used to process the enhancement features through the masked multi-head attention layer of the load prediction model to obtain a temporal feature vector; The prediction unit is used to input the time-series feature vector into the load decoder of the load prediction model to obtain the multivariate load prediction results related to the electrical load, the heat load, and the cooling load output by the load decoder.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the multi-load forecasting method for integrated energy systems as described in any one of claims 1 to 7.

10. A computer program product, the computer program product comprising a computer program, characterized in that, The computer program is read and executed by the processor of the electronic device, causing the electronic device to perform the multi-load forecasting method for integrated energy systems as described in any one of claims 1 to 7.

Images