Comprehensive energy load day-ahead forecasting method based on periodic hash feature-time transformer

Abstract

The application discloses a kind of comprehensive energy load day-ahead forecasting method based on periodical hash feature-time Transformer model, belongs to comprehensive energy system load forecasting technical field, the method includes: the comprehensive energy system load historical data collected is preprocessed;Based on double-layer entropy weight method, the influence factors of comprehensive energy load forecasting are analyzed to determine the input features of the prediction model;Periodical hash feature-time Transformer model is constructed to efficiently extract load periodicity features, cross-energy coupling relationships and their time dependencies;Adaptive conflict projection gradient is used to balance the loss of multiple tasks, dynamically adjust the loss function weights of multiple load forecasting tasks, thereby improving the stability and accuracy of the prediction.

Description

Technical Field

[0001] This invention belongs to the field of integrated energy system load forecasting technology, and in particular relates to a day-ahead forecasting method for integrated energy load based on periodic hash features-time Transformer. Background Technology

[0002] Integrated energy systems, by integrating multiple energy forms, achieve efficient energy utilization and optimized allocation, encompassing multiple stages such as energy production, conversion, storage, and distribution, providing technical support for the efficient operation of energy systems. However, with the increasingly complex coupling relationships between various energy forms, the accuracy of load forecasting faces unprecedented challenges. Although existing day-ahead load forecasting methods can provide some dispatch support, they often struggle to provide high-precision forecasts when dealing with complex multi-energy interactions, the volatility of renewable energy, and the uncertainty of user demand. Therefore, improving the accuracy of day-ahead load forecasting for individual users in integrated energy systems, especially under conditions of multi-energy coupling, has become a key research focus.

[0003] However, in the process of researching and practicing existing technologies, the inventors have found that while some research has been conducted both domestically and internationally on day-ahead forecasting of integrated energy loads, the following issues still require further exploration. First, existing research largely focuses on forecasting single energy loads, paying less attention to the complex coupling relationships between multiple energy loads, and lacks effective modeling methods to capture the nonlinear correlations between various loads. Second, current research on integrated energy load forecasting mostly focuses on ultra-short-term forecasting, while lacking systematic modeling for day-ahead forecasting of multiple energy loads. How to more effectively utilize the characteristics of long-term series to improve the modeling accuracy of day-ahead load forecasting remains a significant challenge. Summary of the Invention

[0004] The purpose of this invention is to provide a comprehensive energy load forecasting method based on the periodic hash feature-time Transformer model, which aims to solve the problems of insufficient characterization of complex coupling relationships, time series dependencies and correlation of external influencing factors among multiple energy loads in the existing technology, thereby improving the accuracy and stability of day-ahead load forecasting of the comprehensive energy system.

[0005] This invention provides a method for establishing a comprehensive energy load forecasting model based on a periodic hash feature-time Transformer model, the method comprising: Preprocess the collected historical load data of the integrated energy system; Correlation analysis of influencing factors in comprehensive energy load forecasting was conducted based on the two-level entropy weight method to determine the input characteristics of the forecasting model. Construct a periodic hash feature-time Transformer model to efficiently extract load periodicity features, cross-energy coupling relationships, and their time dependencies; An adaptive conflict projection gradient multi-task loss equalization method is adopted to dynamically adjust the weights of the loss functions for various load forecasting tasks, thereby improving the stability and accuracy of forecasting.

[0006] Furthermore, the preprocessing of the collected historical load data of the integrated energy system includes: The collected historical data on comprehensive energy load were cleaned and normalized. Furthermore, the correlation analysis of the influencing factors of the comprehensive energy load forecast based on the two-layer entropy weight method is performed to determine the input characteristics of the forecast model, including: The two-layer entropy weight method was used to calculate the comprehensive correlation between each type of load and external influencing factors; Set appropriate thresholds and determine the input features of the prediction model based on the magnitude of the comprehensive correlation index.

[0007] Furthermore, the calculation of the comprehensive correlation between each type of load and external influencing factors using the two-layer entropy weight method includes: The external environmental influencing factors include meteorological factors and time factors; For each external influencing factor, the Pearson correlation coefficient, Spearman correlation coefficient, maximum information coefficient, and grey relational degree are calculated for electrical load, cooling load, and heating load, respectively. The results calculated by each correlation method are normalized to obtain the normalized correlation index. At the level of correlation analysis methods, the entropy weight method is used to calculate the weight of each correlation method and obtain the integrated correlation at the method level. At the load level, the entropy weight method is used again to calculate the weights of different loads, and finally the two-layer entropy weight comprehensive correlation index of each external factor is obtained.

[0008] Furthermore, setting an appropriate threshold and determining the input features of the prediction model based on the magnitude of the comprehensive correlation index includes: The two-layer entropy weighted comprehensive correlation index is compared with a preset threshold. External factors with a comprehensive correlation index greater than the threshold are selected and identified as having a significant impact on the overall energy load. The selected set of significant features is used for training the load forecasting model and subsequent forecasting.

[0009] Furthermore, the construction of the periodic hash feature-time Transformer model to efficiently extract load periodicity features, cross-energy coupling relationships, and their time dependencies includes: The input sequence is converted into a high-dimensional representation through an embedding layer to capture feature information such as time step, feature dimension, and external influencing factors. The input sequence is decomposed into seasonal and trend terms by the sequence decomposition module to enhance the model’s ability to model multidimensional long-term series coupling relationships and complex patterns. Build by N Layer encoder and M The encoder-decoder structure consists of layers of decoders; A periodic feature-temporal multi-head attention mechanism is introduced into the encoder and decoder structure to capture the local and global dependencies of the payload features in time and feature dimensions. In the aforementioned periodic feature-time multi-head attention mechanism, a locality-sensitive hashing mechanism is further introduced to map queries and keys to hash buckets, so as to perform attention computation in a restricted candidate set, thereby reducing computational complexity.

[0010] Furthermore, the step of converting the input sequence into a high-dimensional representation through the embedding layer to capture feature information such as time step, feature dimension, and external influencing factors includes: The token embedding layer is used to extract the dynamic features of the comprehensive energy load. The Conv1d layer performs local convolution operations on the input sequence to capture the change patterns of local features in the sequence. The global timestamp embedding layer is used to capture long-term time patterns, including time information such as month, weekday, date, day type, and hour. The embedding layer maps discrete time information to a continuous vector space to learn the latent features of time. The position embedding layer is used to express the relative relationship between each time position in the sequence. By combining different frequencies of sine and cosine functions, discrete time steps are transformed into continuous vectors, thereby enhancing the expressive power of the time series. The meteorological embedding layer is used to extract meteorological features, including temperature, humidity, and air pressure. The Conv1d layer encodes external influence information, providing the model with key auxiliary information on changes in comprehensive energy load. The embedding layer, through the hierarchical design of the above modules, achieves the fusion of multi-source heterogeneous information and generates a high-dimensional feature representation that can fully reflect the time step, feature dimension and external factors, which is used for the training and subsequent prediction of the load prediction model.

[0011] Furthermore, the step of decomposing the input sequence into seasonal and trend terms through a sequence decomposition module enhances the model's ability to model multidimensional long-term series coupling relationships and complex patterns, including: For the latter half of the model input sequence in the time dimension Perform sequence decomposition to obtain the seasonal term. and periodic terms .

[0012] Specifically, sequence decomposition function The sequence length is kept constant by using a padding operation, and a moving average is performed using the AvgPool function, the formula of which is as follows: In the formula, This represents the kernel size of the moving average function.

[0013] Furthermore, the construction is made by N Layer encoder and M The encoder-decoder structure, composed of layer decoders, includes: Each encoder layer includes a periodic feature-temporal multi-head attention mechanism, a feedforward network, and a sequence decomposition module; Each decoder layer includes a cross-periodic feature-temporal attention mechanism, a feedforward network, and a sequence decomposition module; The final output of the encoder-decoder structure is obtained.

[0014] Furthermore, each layer of the encoder includes a periodic feature-temporal multi-head attention mechanism, a feedforward network, and a sequence decomposition module, comprising: For the input sequence Applying the periodic feature-temporal multi-head attention (PFTA) mechanism, long-term subsequence dependency features are extracted: In the formula, It is the first n Intermediate results from the layer encoder.

[0015] The intermediate representation is used for local feature aggregation via a feedforward (FF) network. Convolutional operations are used to capture short-term similarities within the same period, and sequence decomposition is performed again to obtain the encoder output. In the formula, Indicates the first n The seasonal term of the sequence decomposition module of the layer encoder is used as the first n The output of the layer encoder. This represents the kernel size of the encoder's moving average function. Where: In the formula, It is an activation function. It is the kernel size for convolution operations.

[0016] Furthermore, each layer of the decoder includes a cross-periodic feature-temporal attention mechanism, a feedforward network, and a sequence decomposition module, including: Apply periodic features—temporal multi-head attention (PFTA)—to the seasonal terms of the decoder input: The decoder seasonal term is fused with the encoder output using Cross-Periodic Feature-Temporal Attention (CPFTA): Feature enhancement and sequence decomposition are performed using a feedforward network to obtain the seasonal term output by the m-th layer decoder: The final decoder outputs the seasonal and trend terms as follows: In the formula, They are the first m The intermediate steps of the layer decoder, yes N The output of the layer encoder, It is the first m The seasonal and trend terms output by the layer decoder. It is the size of the kernel of the decoder's moving average function. It is the coefficient corresponding to the trend term of the intermediate step result of the decoder.

[0017] Furthermore, the final output of the encoder-decoder structure includes: The final output of the encoder-decoder structure is formed by a weighted combination of seasonal and trend terms, as shown in the following formula: In the formula, and They are respectively M The output of the layer decoder, is the corresponding coefficient.

[0018] Furthermore, the introduction of a periodic feature-temporal multi-head attention mechanism into the encoder and decoder structure to capture the local and global dependencies of payload features in time and feature dimensions includes: The query Q, key K, and value V are linearly projected h times to form multiple attention heads, and the self-attention function is computed in parallel to obtain the multi-head output: In the formula, , , , and It is a learnable linear projection matrix. The size of each head is... ,in, It is the model dimension. It's the head count.

[0019] Calculate the weights for feature attention and temporal attention separately. Connect the query to the time dimension of the key. It can be reshaped into a two-dimensional structure ,in, Indicates the length of the period. Indicates the number of periodic segments. Each object in the periodic segment... i The temporal attention weight of each attention head can be calculated using the following formula: In the formula, It is the first i The query and key for individual attention spans. , It is the corresponding linear transformation weight matrix. For a period of time, g is a coefficient. This represents the temporal attention weights for all features. To calculate each time step... In the i The feature attention weights of each attention head are calculated using the following formula: In the formula, It is the first i A query for each attention head, key, , It is the corresponding linear transformation weight matrix. This represents the dimension of the key vector. Indicates the first step of all time steps. i Each attention head has its own feature attention weights.

[0020] Through feature attention weights Index the last time dimension to generate a time-effect mapping. This indicates that the first step is performed on each object. m The time impact of the first step, the affected first i The time attention weight of each attention head is defined as: ,exchange The first two dimensions yield all time steps in the th... i Features of individual attention heads – Temporal attention weights .

[0021] The first of each feature Features of individual attention heads – Temporal attention weights With the corresponding value vector Multiply to obtain the corresponding embedded output. .

[0022] The first is obtained by stacking all outputs along the feature dimension. i Embedded output of each attention head This represents the temporal attention weights for all features. The outputs of all attention heads are concatenated to obtain the final output embedding representation. .

[0023] Furthermore, in the aforementioned periodic feature-temporal multi-head attention mechanism, a locality-sensitive hashing mechanism is introduced to map queries and keys to hash buckets, thereby performing attention computation within a restricted candidate set and reducing computational complexity, including: The query and key are bucketed using locality-sensitive hashing, and the data is sorted according to the hash buckets. This is achieved by setting... To enable in-bucket queries Q s and keys K The number of tokens tends to be consistent; if an imbalance still exists, a grouping mechanism is used to sort the input by buckets, and then sort it by token position to obtain a new sequence order. Candidate sets within each bucket Defined as: In the formula, It is the length of the input sequence. It is a query The candidate set contains itself and the elements in the previous bucket. To improve robustness, a multi-round locality-sensitive hashing mechanism is used, with each round using a different hash function, and queries in the same hash bucket from each round are considered separately. Q Take the union of the sets. The hash function for each round is: In the formula, It represents the total number of rounds, and overall, multiple rounds of hashing improve the stability of query and key matching.

[0024] Combining the idea of ​​locality-sensitive hashing, the rewritten attention mechanism is as follows: In the formula, It is a location i The query vector at that location, It is a location j The key vector at that location, It is a locationj The value at that location. is the dimension of the key vector, used to scale the dot product to stabilize the gradient. Indicates position i The set of indexes that a query is interested in uses a locality-sensitive hashing mechanism to map the query and the key to the same hash bucket, thereby reducing the amount of computation. It is a normalization factor used to adjust the attention distribution for different queries. This refers to non-focused locations. j A penalty term is added to ensure that attention computation is performed only within the hash bucket. When using locality-sensitive hashing, all locations are not considered. i Previous elements, but only focus on position i Keys located in the same hash bucket, i.e. ,in, It is a hash function.

[0025] Combining the idea of ​​locality-sensitive hashing, the optimized prediction model output is: In the formula, It is time t The corresponding attention mechanism includes queries and keys. It is a feature f The corresponding attention mechanism includes queries and keys. , , , That is the corresponding weight matrix. For a period of time, g is a coefficient.

[0026] Furthermore, the multi-task loss equalization method employing adaptive conflict projection gradient dynamically adjusts the weights of the loss functions for various load forecasting tasks, thereby improving the stability and accuracy of the forecast, including: The contribution of each load prediction task to the training process is measured to identify the importance of different tasks and guide weight allocation. In multi-task gradient updates, the conflict directions between tasks are coordinated to enhance the collaborative learning effect between tasks. Partial projection is performed on conflicting gradient directions to reduce negative interference and ensure optimization stability; Model parameters are updated based on the coordinated gradient information to achieve stable training and high-precision prediction.

[0027] Furthermore, the measurement of the contribution of each load prediction task during the training process to identify the importance of different tasks and guide weight allocation includes: The volatility and discrimination of the loss of each task during the training process are quantitatively evaluated, thereby adaptively assigning dynamic weights to different tasks to ensure that tasks with higher information gain receive more attention during training.

[0028] Furthermore, the method of coordinating conflict directions among tasks in multi-task gradient updates to enhance the collaborative learning effect among tasks includes: The intensity of conflict is quantitatively characterized by calculating the cosine similarity between task gradients, and EMA and sigmoid are used to map conflict into adaptive tradeoff coefficients. This constrains the lower bound of the similarity between the synthesized gradient and the gradients of each task, reducing the negative impact of conflict on learning. Specifically: It has m One task, t The parameters of the neural network at each time step are , No. k The loss function for each task is: Its gradient is: In the formula, m For the total number of tasks, D For parameter dimensions, For the first k Each task gradient vector.

[0029] The cosine similarity between the gradients of any two tasks is: Define the average negative cosine similarity as the conflict intensity: The degree of conflict is smoothed using the exponential moving average (EMA): In the formula, conflict represents the gradient conflict intensity of the current batch of tasks, and conflict_ema is the smoothed conflict index. This is the EMA smoothing coefficient.

[0030] The conflict level is mapped to (0,1) using the Sigmoid function: The final tradeoff coefficients are obtained through EMA smoothing. : In the formula, This is the output of the conflict response function. ,in, These are the initial tradeoff coefficients after conflict mapping. This represents the minimum allowable range of the weighting factor. This represents the maximum value within the allowable range of the weighting coefficient.

[0031] average gradient Based on this, construct the synthetic gradient

[67] : In the formula, To collaboratively synthesize gradients, The mean of the task gradient. For conflict adaptive trade-off coefficients.

[0032] Furthermore, the partial projection of conflicting gradient directions to reduce negative interference and ensure optimization stability includes: For each task gradient ,like Perform partial projection: In the formula, The projection intensity is 1, where 1 indicates that conflict components are completely eliminated. To prevent extremely small constants with a denominator of 0.

[0033] Furthermore, the step of updating model parameters based on the coordinated gradient information to achieve stable training and high-precision prediction includes: After completing entropy weighting, conflict measurement, synthetic gradient solving, and necessary partial projection correction, the final co-synthetic gradient is used. Update model parameters in the following form: In the formula, For learning rate, These are the updated model parameters.

[0034] An apparatus for establishing a comprehensive energy load day-ahead forecasting model, used to execute the aforementioned comprehensive energy load day-ahead forecasting model establishment method, comprising: The preprocessing module is used to preprocess the collected historical load data of the integrated energy system; The correlation analysis module is used to perform correlation analysis on the influencing factors of comprehensive energy load forecasting based on the two-level entropy weight method, so as to determine the input characteristics of the forecasting model. A prediction model module is established to construct a periodic hash feature-time Transformer model to efficiently extract load periodicity features, cross-energy coupling relationships and their time dependencies; The multi-task loss equalization module is used to dynamically adjust the weights of the loss functions for various load prediction tasks using an adaptive conflict projection gradient multi-task loss equalization method, thereby improving the stability and accuracy of prediction.

[0035] An electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-described method for establishing a day-ahead forecasting model for integrated energy load.

[0036] A computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described method for establishing a day-ahead forecasting model for integrated energy load.

[0037] Compared with existing technologies, this invention fully considers the periodicity of load, multi-timescale characteristics, and complex coupling relationships among multiple energy types in the day-ahead forecasting model of integrated energy load. By introducing a two-layer entropy weight correlation analysis method, the relative importance of different influencing factors in integrated energy load forecasting is effectively quantified, ensuring the scientific nature and representativeness of the input features. Combining a periodic hash feature-time Transformer model, the periodic feature extraction module is used to mine the periodic variation patterns of load, and the long-term and short-term dependencies of load in the time dimension are modeled through a time Transformer structure, thereby achieving collaborative learning of periodic patterns and time dynamics. Simultaneously, an adaptive conflict projection gradient multi-task loss balancing mechanism is introduced, which can dynamically adjust the loss weights among multiple load forecasting tasks, alleviating task conflicts and improving overall forecast stability and generalization performance. Through the joint design of periodic feature perception, time-dependent modeling, and multi-task adaptive optimization, this invention significantly improves the accuracy and robustness of integrated energy load forecasting, possessing stronger model generalization ability and engineering application value. Attached Figure Description

[0038] Figure 1 This is a flowchart of the method for establishing a comprehensive energy load day-ahead forecasting model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the device for establishing a comprehensive energy load day-ahead forecasting model according to Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the device for establishing a day-ahead prediction model for the integrated energy load in Embodiment 2 of the present invention. Detailed Implementation

[0039] To overcome the shortcomings of existing technologies, this invention proposes a comprehensive energy system load forecasting method based on a periodic hash feature-time Transformer model. It analyzes multiple influencing factors of comprehensive energy load forecasting and establishes a day-ahead forecasting model for comprehensive energy load.

[0040] The technical solutions of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0041] This invention provides a method for predicting day-ahead energy load based on a periodic hash feature-time Transformer model. The flowchart of the method is shown below. Figure 1 As shown, the specific processing includes the following: Step 1 (S101) involves preprocessing and dimensionality reduction of the collected historical day-ahead comprehensive energy load data. Specifically, the following processing methods can be employed: (1) Conduct load data preprocessing, remove outliers and complete or remove missing records; (2) Normalize the features of each dimension to ensure that the scale of the input features is consistent.

[0042] Step 2 (S102): Based on the two-layer entropy weight method, conduct correlation analysis on the influencing factors of comprehensive energy load forecasting to determine the input characteristics of the forecasting model. Specifically, the following processing can be adopted: Step 2.1: The comprehensive correlation between each type of load and external influencing factors is calculated using the two-layer entropy weight method. The specific process is as follows: The external environmental influencing factors include meteorological factors and time factors; For each external influencing factor, the Pearson correlation coefficient, Spearman correlation coefficient, maximum information coefficient, and grey relational degree are calculated for electrical load, cooling load, and heating load, respectively. The results calculated by each correlation method are normalized to obtain the normalized correlation index. At the level of correlation analysis methods, the entropy weight method is used to calculate the weight of each correlation method and obtain the integrated correlation at the method level. At the load level, the entropy weight method is used again to calculate the weights of different loads, and finally the two-layer entropy weight comprehensive correlation index of each external factor is obtained.

[0043] Step 2.2: Set an appropriate threshold and determine the input features of the prediction model based on the magnitude of the comprehensive correlation index. The specific process is as follows: The two-layer entropy weighted comprehensive correlation index is compared with a preset threshold. External factors with a comprehensive correlation index greater than the threshold are selected and identified as having a significant impact on the overall energy load. The selected set of significant features is used for training the load forecasting model and subsequent forecasting.

[0044] Step 3 (S103): Construct a periodic hash feature-time Transformer model to efficiently extract load periodicity features, cross-energy coupling relationships, and their time dependencies. Specifically, the following processing can be adopted: Step 3.1: The input sequence is converted into a high-dimensional representation through an embedding layer to capture feature information such as time step, feature dimension, and external influencing factors. The specific process is as follows: The token embedding layer is used to extract the dynamic features of the comprehensive energy load. The Conv1d layer performs local convolution operations on the input sequence to capture the change patterns of local features in the sequence. The global timestamp embedding layer is used to capture long-term time patterns, including time information such as month, weekday, date, day type, and hour. The embedding layer maps discrete time information to a continuous vector space to learn the latent features of time. The position embedding layer is used to express the relative relationship between each time position in the sequence. By combining different frequencies of sine and cosine functions, discrete time steps are transformed into continuous vectors, thereby enhancing the expressive power of the time series. The meteorological embedding layer is used to extract meteorological features, including temperature, humidity, and air pressure. The Conv1d layer encodes external influence information, providing the model with key auxiliary information on changes in comprehensive energy load. The embedding layer, through the hierarchical design of the above modules, achieves the fusion of multi-source heterogeneous information and generates a high-dimensional feature representation that can fully reflect the time step, feature dimension and external factors, which is used for the training and subsequent prediction of the load prediction model.

[0045] Step 3.2 involves decomposing the input sequence into seasonal and trend terms using the sequence decomposition module to enhance the model's ability to model multidimensional long-term series coupling relationships and complex patterns. The specific process is as follows: For the latter half of the model input sequence in the time dimension Perform sequence decomposition to obtain the seasonal term. and periodic terms .

[0046] Specifically, sequence decomposition function The sequence length is kept constant by using a padding operation, and a moving average is performed using the AvgPool function, the formula of which is as follows: In the formula, This represents the kernel size of the moving average function.

[0047] Step 3.3, constructing from N Layer encoder and M The encoder-decoder structure, composed of layers and decoders, is described in the following process: Step 3.3.1: Each encoder layer includes a periodic feature-temporal multi-head attention mechanism, a feedforward network, and a sequence decomposition module. The specific process is as follows: For the input sequence Applying the periodic feature-temporal multi-head attention (PFTA) mechanism, long-term subsequence dependency features are extracted: In the formula, It is the first n Intermediate results from the layer encoder.

[0048] The intermediate representation is used for local feature aggregation via a feedforward (FF) network. Convolutional operations are used to capture short-term similarities within the same period, and sequence decomposition is performed again to obtain the encoder output. In the formula, Indicates the first n The seasonal term of the sequence decomposition module of the layer encoder is used as the first n The output of the layer encoder. This represents the kernel size of the encoder's moving average function. Where: In the formula, It is an activation function. It is the kernel size for convolution operations.

[0049] Step 3.3.2: Each decoder layer includes a cross-periodic feature-temporal attention mechanism, a feedforward network, and a sequence decomposition module. The specific process is as follows: Apply periodic features—temporal multi-head attention (PFTA)—to the seasonal terms of the decoder input: The decoder seasonal term is fused with the encoder output using Cross-Periodic Feature-Temporal Attention (CPFTA): Feature enhancement and sequence decomposition are performed using a feedforward network to obtain the seasonal term output by the m-th layer decoder: The final decoder outputs the seasonal and trend terms as follows: In the formula, They are the first m The intermediate steps of the layer decoder, yes N The output of the layer encoder, It is the first m The seasonal and trend terms output by the layer decoder. It is the size of the kernel of the decoder's moving average function. It is the coefficient corresponding to the trend term of the intermediate step result of the decoder.

[0050] Step 3.3.3 yields the final output of the encoder-decoder structure, the specific process of which is as follows: The final output of the encoder-decoder structure is formed by a weighted combination of seasonal and trend terms, as shown in the following formula: In the formula, and They are respectively M The output of the layer decoder, is the corresponding coefficient.

[0051] Step 3.4 introduces a periodic feature-temporal multi-head attention mechanism into the encoder and decoder structure to capture the local and global dependencies of the payload features in time and feature dimensions. The specific process is as follows: The query Q, key K, and value V are linearly projected h times to form multiple attention heads, and the self-attention function is computed in parallel to obtain the multi-head output: In the formula, , , , and It is a learnable linear projection matrix. The size of each head is... ,in, It is the model dimension. It's the head count.

[0052] Calculate the weights for feature attention and temporal attention separately. Connect the query to the time dimension of the key. It can be reshaped into a two-dimensional structure ,in, Indicates the length of the period. Indicates the number of periodic segments. Each object in the periodic segment... i The temporal attention weight of each attention head can be calculated using the following formula: In the formula, It is the first i The query and key for individual attention spans. , It is the corresponding linear transformation weight matrix. For a period of time, g is a coefficient. This represents the temporal attention weights for all features. To calculate each time step... In the i The feature attention weights of each attention head are calculated using the following formula: In the formula, It is the first i A query for each attention head, key, , It is the corresponding linear transformation weight matrix. This represents the dimension of the key vector. Indicates the first step of all time steps. i Each attention head has its own feature attention weights.

[0053] Through feature attention weights Index the last time dimension to generate a time-effect mapping. This indicates that the first step is performed on each object. m The time impact of the first step, the affected first i The time attention weight of each attention head is defined as: ,exchange The first two dimensions yield all time steps in the th... i Features of individual attention heads – Temporal attention weights .

[0054] The first of each feature Features of individual attention heads – Temporal attention weights With the corresponding value vector Multiply to obtain the corresponding embedded output. .

[0055] The first is obtained by stacking all outputs along the feature dimension. i Embedded output of each attention head This represents the temporal attention weights for all features. The outputs of all attention heads are concatenated to obtain the final output embedding representation: .

[0056] Step 3.5: In the periodic feature-temporal multi-head attention mechanism, a locality-sensitive hashing mechanism is further introduced to map queries and keys to hash buckets, so as to perform attention calculations in a restricted candidate set, thereby reducing computational complexity. The specific process is as follows: The query and key are bucketed using locality-sensitive hashing, and the data is sorted according to the hash buckets. This is achieved by setting... To enable in-bucket queries Q s and keysK The number of tokens tends to be consistent; if an imbalance still exists, a grouping mechanism is used to sort the input by buckets, and then sort it by token position to obtain a new sequence order. Candidate sets within each bucket Defined as: In the formula, It is the length of the input sequence. It is a query The candidate set contains itself and the elements in the previous bucket. To improve robustness, a multi-round locality-sensitive hashing mechanism is used, with each round using a different hash function, and queries in the same hash bucket from each round are considered separately. Q Take the union of the sets. The hash function for each round is: In the formula, It represents the total number of rounds, and overall, multiple rounds of hashing improve the stability of query and key matching.

[0057] Combining the idea of ​​locality-sensitive hashing, the rewritten attention mechanism is as follows: In the formula, It is a location i The query vector at that location, It is a location j The key vector at that location, It is a location j The value at that location. is the dimension of the key vector, used to scale the dot product to stabilize the gradient. Indicates position i The set of indexes that a query is interested in uses a locality-sensitive hashing mechanism to map the query and the key to the same hash bucket, thereby reducing the amount of computation. It is a normalization factor used to adjust the attention distribution for different queries. This refers to non-focused locations. j A penalty term is added to ensure that attention computation is performed only within the hash bucket. When using locality-sensitive hashing, all locations are not considered. i Previous elements, but only focus on position i Keys located in the same hash bucket, i.e. ,in, It is a hash function.

[0058] Combining the idea of ​​locality-sensitive hashing, the optimized prediction model output is: In the formula, It is timet The corresponding attention mechanism includes queries and keys. It is a feature f The corresponding attention mechanism includes queries and keys. , , , That is the corresponding weight matrix. For a period of time, g is a coefficient.

[0059] Step 4 (S104) employs an adaptive conflict projection gradient multi-task loss balancing method to dynamically adjust the weights of the loss functions for various load prediction tasks, thereby improving the stability and accuracy of the prediction. Specifically, the following processing can be implemented: Step 4.1: Measure the contribution of each load prediction task during the training process to identify the importance of different tasks and guide weight allocation. The specific process is as follows: The volatility and discrimination of the loss of each task during the training process are quantitatively evaluated, thereby adaptively assigning dynamic weights to different tasks to ensure that tasks with higher information gain receive more attention during training.

[0060] Step 4.2: In multi-task gradient updates, coordinate the conflict directions between tasks to enhance the collaborative learning effect between tasks. The specific process is as follows: The intensity of conflict is quantitatively characterized by calculating the cosine similarity between task gradients, and EMA and sigmoid are used to map conflict into adaptive tradeoff coefficients. This constrains the lower bound of the similarity between the synthesized gradient and the gradients of each task, reducing the negative impact of conflict on learning. Specifically: It has m One task, t The parameters of the neural network at each time step are , No. k The loss function for each task is: Its gradient is: In the formula, m For the total number of tasks, D For parameter dimensions, For the first k Each task gradient vector.

[0061] The cosine similarity between the gradients of any two tasks is: Define the average negative cosine similarity as the conflict intensity: The degree of conflict is smoothed using the exponential moving average (EMA): In the formula, conflict represents the gradient conflict intensity of the current batch of tasks, and conflict_ema is the smoothed conflict index. This is the EMA smoothing coefficient.

[0062] The conflict level is mapped to (0,1) using the Sigmoid function: The final tradeoff coefficients are obtained through EMA smoothing. : In the formula, This is the output of the conflict response function. ,in, These are the initial tradeoff coefficients after conflict mapping. This represents the minimum allowable range of the weighting factor. This represents the maximum value within the allowable range of the weighting coefficient.

[0063] average gradient Based on this, construct the synthetic gradient: In the formula, To collaboratively synthesize gradients, The mean of the task gradient. For conflict adaptive trade-off coefficients.

[0064] Step 4.3 involves partially projecting conflicting gradient directions to reduce negative interference and ensure optimization stability. The specific process is as follows: For each task gradient ,like Perform partial projection: In the formula, The projection intensity is 1, where 1 indicates that conflict components are completely eliminated. To prevent extremely small constants with a denominator of 0.

[0065] Step 4.4 involves updating the model parameters based on the coordinated gradient information to achieve stable training and high-precision prediction. The specific process is as follows: After completing entropy weighting, conflict measurement, synthetic gradient solving, and necessary partial projection correction, the final co-synthetic gradient is used. Update model parameters in the following form: In the formula, For learning rate, These are the updated model parameters.

[0066] In other words, the specific process for the day-ahead deterministic forecasting of integrated energy load is as follows: (1) Data processing and prediction process The experimental dataset used in this embodiment originates from historical real-world operational data of a typical building on the Tempe campus of Arizona State University (ASU). This building serves as a standard integrated building energy system, and its dataset covers electrical load (electricity consumption of internal electrical equipment), heat load (thermal energy demand of steam or hot water systems), and cooling load (cooling demand of chilled water systems). Additionally, outdoor meteorological data for the region provided by the U.S. National Climate Data Center were simultaneously acquired. The time span was selected from 01:00 on January 1, 2017 to 23:00 on January 31, 2021, with a sampling frequency of 1 hour. Before model training, to overcome the interference caused by the randomness and volatility of electrical, cooling, and heating loads, as well as potential missing values, in the original data, linear interpolation was first used to fill in the missing records. Secondly, to eliminate the differences in dimensions and magnitudes between different load characteristics, accelerate gradient convergence, and prevent gradient explosion, the extreme value normalization method was used to map all input variables to the [0,1] interval. The normalization calculation formula is as follows: In the formula, This represents the normalized data. The original input data, and These represent the maximum and minimum values ​​of the feature variable in the sample set, respectively. The processed dataset is divided into training, validation, and test sets in a 6:2:2 ratio. All deep learning models in this embodiment are built on the PyTorch framework and trained on a hardware platform equipped with an NVIDIA GeForce RTX 3080Ti GPU.

[0067] (2) Predictive performance evaluation indicators To objectively quantify and compare the point prediction performance of different models, this study uses root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators. The formulas for calculating RMSE and MAE are as follows: In the formula, For the first Load forecast values ​​for each sample The actual load value observed. Represents the total number of samples. RMSE, calculated by taking the square root of the mean of the squared errors, can effectively amplify larger deviations in predictions; MAE, on the other hand, reflects the average level of the absolute error between the predicted and actual values. The smaller the values ​​of RMSE and MAE, the lower the model's prediction error and the higher its accuracy; while R... 2 The closer the value is to 1, the better the model fits the actual load curve.

[0068] (3) Comparative experimental analysis To verify the superiority of the proposed method (denoted as PHFTformer), this embodiment selects some mainstream and cutting-edge prediction models for horizontal comparative analysis. The baseline models for comparison include: Bidirectional Long Short-Term Memory Network (BiLSTM), Convolutional Bidirectional Gated Network with Integrating Attention Mechanism (CNN-BiGRU-ATT), Long Sequence Temporal Optimization Model (Informer), Autocorrelation-based Model (Autoformer), and Feature-Time Attention Model (FTformer). The evaluation results of the day-ahead comprehensive energy (electricity, cooling, and heating) load predicted by each model are shown in Table 1.

[0069] Table 1 As can be seen from the experimental data analysis in Table 1, the RMSE and MAE indices of the proposed model are lower than those of the comparative model in the prediction of electricity, cooling and heating loads, which fully demonstrates the advancement and effectiveness of the proposed model in multi-energy coupling feature extraction and high-precision day-ahead prediction tasks.

[0070] Device Examples According to embodiments of the present invention, an apparatus for establishing a comprehensive energy load day-ahead forecasting model is provided, such as... Figure 2 As shown, the integrated energy load day-ahead forecasting model establishment device according to an embodiment of the present invention specifically includes: The preprocessing module 60 is used to preprocess the collected historical load data of the integrated energy system.

[0071] The correlation analysis module 62 is used to perform correlation analysis on the influencing factors of comprehensive energy load forecasting based on the two-layer entropy weight method, so as to determine the input characteristics of the forecasting model. A prediction model module 64 is established to construct a periodic hash feature-time Transformer model to efficiently extract load periodicity features, cross-energy coupling relationships and their time dependencies. The multi-task loss equalization module 66 is used to dynamically adjust the weights of the loss functions of various load prediction tasks by adopting the multi-task loss equalization method of adaptive conflict projection gradient, thereby improving the stability and accuracy of prediction.

[0072] The embodiments of the present invention are device embodiments corresponding to the above method embodiments. The specific operation of each module can be understood with reference to the description of the method embodiments, and will not be repeated here.

[0073] This invention provides an electronic device, such as... Figure 3 As shown, it includes: a memory 70, a processor 72, and a computer program stored in the memory 70 and executable on the processor 72, wherein the computer program, when executed by the processor 72, performs the steps as described in the method embodiment.

[0074] This invention provides a computer-readable storage medium storing an information transmission implementation program. When executed by a processor 72, the program performs the steps described in the method embodiment. The computer-readable storage medium in this embodiment includes, but is not limited to, ROM, RAM, magnetic disk, or optical disk.

[0075] The above description is merely an embodiment of this document and is not intended to limit the scope of this document. Various modifications and variations can be made to this document by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this document should be included within the scope of the claims of this document.

Claims

1. A comprehensive day-ahead energy load forecasting method based on periodic hash features-time Transformer, characterized in that, include: S1 preprocesses the collected historical load data of the integrated energy system; S2 uses the two-layer entropy weight method to conduct correlation analysis on the influencing factors of comprehensive energy load forecasting in order to determine the input characteristics of the forecasting model; S3 constructs a periodic hash feature-time Transformer model to efficiently extract load periodic features, cross-energy coupling relationships, and their time dependencies; S4 employs an adaptive conflict projection gradient multi-task loss equalization method to dynamically adjust the weights of the loss functions for various load forecasting tasks, thereby improving the stability and accuracy of forecasting.

2. The integrated day-ahead energy load forecasting method based on periodic hash features-time Transformer according to claim 1, characterized in that, Step S2 includes: S201 uses the two-layer entropy weight method to calculate the comprehensive correlation between each type of load and external influencing factors. The external environmental influencing factors include meteorological elements and time elements. For each external influencing factor, the Pearson correlation coefficient, Spearman correlation coefficient, maximum information coefficient and grey relational degree are calculated with electrical load, cooling load and heating load respectively. The results of each correlation method are normalized to obtain a normalized correlation index. At the level of correlation analysis methods, the entropy weight method is used to calculate the weight of each correlation method and to obtain the integrated correlation at the method level. At the load level, the entropy weight method is used again to calculate the weights of different loads, and finally the two-layer entropy weight comprehensive correlation index of each external factor is obtained. S202 sets an appropriate threshold and determines the input features of the prediction model based on the magnitude of the comprehensive correlation index, including: The two-layer entropy weighted comprehensive correlation index is compared with a preset threshold. External factors with a comprehensive correlation index greater than the threshold are selected and identified as having a significant impact on the overall energy load. The selected set of significant features is used for training the load forecasting model and subsequent forecasting.

3. The integrated day-ahead energy load forecasting method based on periodic hash features-time Transformer according to claim 1, characterized in that, S3 include: S301 converts the input sequence into a high-dimensional representation through an embedding layer to capture feature information such as time step, feature dimension, and external influencing factors. Specifically: The token embedding layer is used to extract the dynamic features of the comprehensive energy load. The Conv1d layer performs local convolution operations on the input sequence to capture the change patterns of local features in the sequence. A global timestamp embedding layer is used to capture long-term time patterns, including time information such as month, weekday, date, day type, and hour. The embedding layer maps discrete time information to a continuous vector space to learn the latent features of time. The position embedding layer is used to express the relative relationship between each time position in the sequence. By combining different frequencies of sine and cosine functions, discrete time steps are transformed into continuous vectors, thereby enhancing the expressive power of the time series. The meteorological embedding layer is used to extract meteorological features, including temperature, humidity, and air pressure. The Conv1d layer encodes external influence information, providing the model with key auxiliary information on changes in comprehensive energy load. The embedding layer, through the hierarchical design of the above modules, achieves the fusion of multi-source heterogeneous information and generates a high-dimensional feature representation that can fully reflect the time step, feature dimension and external factors, which is used for the training and subsequent prediction of the load prediction model. S302 uses a sequence decomposition module to decompose the input sequence into seasonal and trend terms, enhancing the model's ability to model multidimensional long-term series coupling relationships and complex patterns, including: For the latter half of the model input sequence in the time dimension Perform sequence decomposition to obtain the seasonal term. and periodic terms , ， Specifically, sequence decomposition function The sequence length is kept constant by using a padding operation, and a moving average is performed using the AvgPool function, the formula of which is as follows: ， ， In the formula, This represents the kernel size of the moving average function.

4. The integrated day-ahead energy load forecasting method based on periodic hash features-time Transformer according to claim 3, characterized in that, S3 include: S303 is constructed from N Layer encoder and M The encoder-decoder structure, composed of layer decoders, includes: Each encoder layer includes a periodic feature-temporal multi-head attention mechanism, a feedforward network, and a sequence decomposition module; Each decoder layer includes a cross-periodic feature-temporal attention mechanism, a feedforward network, and a sequence decomposition module; The final output of the encoder-decoder structure is obtained; Each layer of the S35 encoder includes a periodic feature-temporal multi-head attention mechanism, a feedforward network, and a sequence decomposition module, including: For the input sequence Applying the periodic feature-temporal multi-head attention (PFTA) mechanism, long-term subsequence dependency features are extracted: ， In the formula, It is the first n Intermediate results from the layer encoder; The intermediate representation is used for local feature aggregation via a feedforward (FF) network. Convolutional operations are used to capture short-term similarities within the same period, and sequence decomposition is performed again to obtain the encoder output. ， In the formula, Indicates the first n The seasonal term of the sequence decomposition module of the layer encoder is used as the first n The output of the layer encoder, Let be the size of the kernel of the encoder's moving average function, where: , In the formula, It is an activation function. It is the kernel size for convolution operations.

5. The integrated day-ahead energy load forecasting method based on periodic hash feature-time Transformer according to claim 4, characterized in that, Each decoder layer includes a cross-periodic feature-temporal attention mechanism, a feedforward network, and a sequence decomposition module, including: Apply periodic features—temporal multi-head attention (PFTA)—to the seasonal terms of the decoder input: ， The decoder seasonal term is fused with the encoder output using Cross-Periodic Feature-Temporal Attention (CPFTA): ， Feature enhancement and sequence decomposition are performed using a feedforward network to obtain the seasonal term output by the m-th layer decoder: ， The final decoder outputs the seasonal and trend terms as follows: ， In the formula, They are the first m The intermediate steps of the layer decoder, yes N The output of the layer encoder, It is the first m The seasonal and trend terms output by the layer decoder It is the size of the kernel of the decoder's moving average function. These are the coefficients corresponding to the trend term in the intermediate steps of the decoder; The final output of the encoder-decoder structure is obtained by weighting and combining seasonal and trend terms, as shown in the following formula: ， In the formula, and They are respectively M The output of the layer decoder, For the corresponding coefficients; A periodic feature-temporal multi-head attention mechanism is introduced into the encoder and decoder architecture to capture the local and global dependencies of payload features in time and feature dimensions, including: The query Q, key K, and value V are linearly projected h times to form multiple attention heads, and the self-attention function is computed in parallel to obtain the multi-head output: ， In the formula, , , , and It is a learnable linear projection matrix, with each head having a size of [value missing]. ,in, It is the model dimension. It is the number of heads; Calculate the weights of feature attention and temporal attention separately, and link the query to the time dimension of the key. It can be reshaped into a two-dimensional structure ,in, Indicates the length of the period. Indicates the number of periodic segments, each object in the 1st... i The temporal attention weight of each attention head can be calculated using the following formula: ，， In the formula, It is the first i The query and key for individual attention spans. , It is the corresponding linear transformation weight matrix. For a period of time, g For coefficients, This represents the temporal attention weights for all features, in order to compute the values ​​at each time step. In the i The feature attention weights of each attention head are calculated using the following formula: ， In the formula, It is the first i A query for each attention head, key, , It is the corresponding linear transformation weight matrix. The dimension of the key vector. Indicates the first step of all time steps. i Each attention head has its own feature attention weights; Through feature attention weights Index the last time dimension to generate a time-effect mapping. This indicates that the first step is performed on each object. m The time impact of the first step, the affected first i The time attention weight of each attention head is defined as: ， ,exchange The first two dimensions yield all time steps in the th... i Features of individual attention heads – Temporal attention weights ; The first of each feature Features of individual attention heads – Temporal attention weights With the corresponding value vector Multiply to obtain the corresponding embedded output. ; ， The first is obtained by stacking all outputs along the feature dimension. i Embedded output of each attention head Representing the temporal attention weights of all features, the outputs of all attention heads are concatenated to obtain the final output embedding representation: 。 6. The integrated day-ahead energy load forecasting method based on periodic hash features-time Transformer according to claim 5, characterized in that, S3 include: S305 further introduces a locality-sensitive hashing mechanism into the periodic feature-time multi-head attention mechanism to map queries and keys to hash buckets, thereby performing attention computation within a restricted candidate set and reducing computational complexity, including: The query and key are bucketed using locality-sensitive hashing, and the data is sorted according to the hash buckets. This is achieved by setting... To enable in-bucket queries Q s and keys K The number of tokens tends to be consistent; if an imbalance still exists, a grouping mechanism is used to sort the input by buckets, and then sort it by token position to obtain a new sequence order. Candidate sets within each bucket Defined as: ， In the formula, It is the length of the input sequence. It is a query The candidate set contains the element itself and the elements in the previous bucket. To improve robustness, a multi-round locality-sensitive hashing mechanism is used, with each round using a different hash function, and queries in the same hash bucket in each round are considered separately. Q Taking the union of sets, the hash function for each round is: ， In the formula, It represents the total number of rounds; overall, multiple rounds of hashing improve the stability of query and key matching. Combining the idea of ​​locality-sensitive hashing, the rewritten attention mechanism is as follows: ， In the formula, It is a location i The query vector at that location, It is a location j The key vector at that location, It is a location j The value at that location, is the dimension of the key vector, used to scale the dot product to stabilize the gradient. Indicates position i The set of indexes that a query focuses on uses a locality-sensitive hashing mechanism to map queries and keys to the same hash bucket, thereby reducing computational overhead. It is a normalization factor used to adjust the attention distribution for different queries. This refers to non-focused locations. j The penalty term is used to ensure that attention computation is performed only within the hash bucket, whereas when using locality-sensitive hashing, not all locations are considered. i Previous elements, but only focus on position i Keys located in the same hash bucket, i.e. ,in, It is a hash function; Combining the idea of ​​locality-sensitive hashing, the optimized prediction model output is: ， In the formula, It is time t The corresponding attention mechanism includes queries and keys. It is a feature f The corresponding attention mechanism includes queries and keys. , , , That is the corresponding weight matrix. For a period of time, g is a coefficient.

7. The integrated day-ahead energy load forecasting method based on periodic hash feature-time Transformer according to claim 1, characterized in that, S4 include: S401 measures the contribution of each load prediction task during the training process to identify the importance of different tasks and guide weight allocation, including: quantifying the variability and discrimination of the loss of each task during the training process, thereby adaptively assigning dynamic weights to different tasks to ensure that tasks with higher information gain receive more attention during training. S402 coordinates the conflict directions between tasks in multi-task gradient updates to enhance the collaborative learning effect. This includes: quantitatively characterizing the conflict intensity by calculating the cosine similarity between task gradients, and using EMA and Sigmoid to map the conflict into adaptive tradeoff coefficients, thereby constraining the lower bound of the similarity between the synthesized gradient and the gradients of each task, reducing the negative impact of conflict on learning. Specifically: It has m One task, t The parameters of the neural network at each time step are , No. k The loss function for each task is: Its gradient is: ， In the formula, m The total number of tasks. D For parameter dimensions, For the first k Each task gradient vector; The cosine similarity between the gradients of any two tasks is: ， Define the average negative cosine similarity as the conflict intensity: ， The degree of conflict is smoothed using the exponential moving average (EMA): ， In the formula, conflict represents the gradient conflict intensity of the current batch of tasks, and conflict_ema is the smoothed conflict index. The EMA smoothing coefficient; The conflict level is mapped to (0,1) using the Sigmoid function: ， The final tradeoff coefficients are obtained through EMA smoothing. : ， In the formula, For the output of the conflict response function, ,in, These are the initial tradeoff coefficients after conflict mapping. This represents the minimum allowable range of the weighting factor. This represents the maximum value within the allowable range of the tradeoff coefficients; average gradient Based on this, construct the synthetic gradient [67]: ， In the formula, To collaboratively synthesize gradients, The mean of the task gradient. For conflict adaptive trade-off coefficients; S403 partially projects conflicting gradient directions to reduce negative interference and ensure optimization stability, including: For each task gradient ,like Perform partial projection: ， In the formula, The projection intensity is 1, where 1 indicates complete elimination of conflict components. To prevent extremely small constants with a denominator of 0; S404 updates model parameters based on the coordinated gradient information, achieving stable training and high-precision prediction, including: After completing entropy weighting, conflict measurement, synthetic gradient solving, and necessary partial projection correction, the final co-synthetic gradient is used. Update model parameters in the following form: ， In the formula, For learning rate, These are the updated model parameters.

8. A device for establishing a comprehensive day-ahead energy load forecasting model based on periodic hash features-time Transformer, characterized in that, A method for establishing a comprehensive energy load day-ahead forecasting model as described in any one of claims 1 to 7, comprising: The preprocessing module is used to preprocess the collected historical load data of the integrated energy system; The correlation analysis module is used to perform correlation analysis on the influencing factors of comprehensive energy load forecasting based on the two-level entropy weight method, so as to determine the input characteristics of the forecasting model. A prediction model module is established to construct a periodic hash feature-time Transformer model to efficiently extract load periodicity features, cross-energy coupling relationships and their time dependencies; The multi-task loss equalization module is used to dynamically adjust the weights of the loss functions for various load forecasting tasks using an adaptive conflict projection gradient multi-task loss equalization method, thereby improving the stability and accuracy of forecasting.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program is executed by the processor to implement the steps of the method for establishing an integrated energy load day-ahead forecasting model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the integrated energy load day-ahead forecasting model establishment method as described in any one of claims 1 to 7.

Images