Comprehensive energy system multi-element load prediction method based on double-layer decomposition and reconstruction and TECNFormer

Abstract

The invention provides a comprehensive energy system multi-element load prediction method based on double-layer decomposition and reconstruction and TECNFormer, and the method comprises the steps: obtaining standardized input data, meteorological factors and time characteristics, which are obtained through the preprocessing of a historical multi-energy load sequence of a comprehensive energy system, wherein the historical multi-energy load sequence comprises an electric load sequence, a cold load sequence and a heat load sequence; performing double-layer modal decomposition on the standardized input data, and classifying the standardized input data into different types of dynamic components; respectively inputting the dynamic components into a time sequence enhanced convolution module, and extracting multi-scale local features by combining multiple heterogeneous differential convolution operators with causal convolution; taking a TECNFormer composed of a time sequence enhanced convolution module and an improved long sequence prediction network as a unified shared feature extraction layer, combining the improved long sequence prediction network with a bidirectional long-short-term memory network and a sparse attention mechanism to capture long-range dependence and local details, and obtaining joint modeling features; on the basis of a hard shared network architecture, a multi-task branch is arranged at an output end, and electric, cold and heat load prediction results are synchronously output.

Description

Technical Field

[0001] This invention belongs to the field of integrated energy system operation optimization and load forecasting technology, specifically involving a multi-dimensional load forecasting method for integrated energy systems based on two-layer decomposition and reconstruction and TECNFormer. Background Technology

[0002] Electricity load forecasting has evolved along multiple technical routes, from statistical learning and machine learning to deep learning. Traditional statistical and machine learning methods are mainly used for short-term user forecasting scenarios, but their prediction performance is poor when faced with load sequences that are strongly non-stationary, multi-scale coupled, and have a large number of exogenous disturbances. Deep learning is gradually becoming mainstream, with recurrent networks and convolutional networks being widely used, and Transformers for long-sequence requirements are also developing rapidly. At the same time, combined methods of decomposition and reconstruction frameworks and neural networks are also widely adopted, that is, first decomposing the signal into different modes, and then modeling and reconstructing them separately to improve accuracy. With the rise of integrated energy systems, the need for joint modeling of multi-energy loads (electricity / cooling / heating, etc.) and their cross-spatial correlations is becoming increasingly prominent.

[0003] Overall, existing technologies still have the following shortcomings: First, the handling of non-stationarity and multi-scale coupling is insufficient. Classical decomposition methods are prone to mode aliasing, and while improved methods can suppress noise, they are sensitive to parameters, easily leading to fluctuations in decomposition quality. Second, it is still difficult to balance global dependency acquisition and computational efficiency under long prediction windows. Although Transformer-type methods introduce sparse attention, autocorrelation, or frequency domain mapping, the problems of effective inductive bias and dataset sensitivity have not been solved. Third, the pipelined prediction process of decomposition and reconstruction is prone to error propagation and parameter dependence, and preprocessing decomposition is difficult to express the hierarchical interaction of future periods. Fourth, cross-component / cross-energy correlations are not fully utilized, and there is a lack of deterministic analysis, cross-component constraints, and information sharing mechanisms, making it difficult to suppress spurious correlations and duplicate information. The time-varying commonalities among multi-energy loads have not been systematically and jointly modeled. Fifth, interpretability and physical consistency are poor. Deep models are mostly "black boxes," lacking structured explanations for the contributions of trends, seasons, and event impacts, affecting the credibility of engineering implementation. Summary of the Invention

[0004] To address the shortcomings of existing technologies in load forecasting of integrated energy systems, such as insufficient handling of non-stationarity and multi-scale coupling, difficulty in balancing long-range dependency capture and computational efficiency, inadequate utilization of cross-energy component correlations, and error propagation, this invention provides a multi-element load forecasting method for integrated energy systems based on two-layer decomposition and reconstruction and TECNFormer. This method aims to solve problems such as modal aliasing of strongly non-stationary load sequences, low efficiency of long-range dependency modeling, and insufficient utilization of cross-energy correlations.

[0005] This method first decouples the load sequence through a two-layer mode decomposition framework: the first layer uses robust empirical mode decomposition to obtain intrinsic mode components and residuals, and optimizes the decomposition stability through mirror symmetric extension and iterative sieving; the second layer uses continuous variational mode decomposition to successively subdivide the initial component set, and obtains fine-grained candidate components by combining frequency separation precision control. Based on a multi-index fusion mechanism of frequency information (Hilbert transform time-frequency analysis), correlation degree (weighted fusion of autocorrelation coefficient and maximum information coefficient), and complexity (multi-scale permutation entropy), the candidate components are adaptively classified into high, medium, and low dynamic components, improving the physical interpretability of the components.

[0006] For the classified dynamic components, a temporal enhancement convolution module containing five types of heterogeneous differential convolution operators (center / rotation / edge / horizontal / vertical differential convolution) is designed. This module is combined with a temporally dilated convolution block with an exponential inflation rate of 2 and causal convolution to extract multi-scale local features to suppress non-stationary noise spurious correlations. The features are then fused with adaptive positional encoding and temporal encoding containing learnable periodic parameters and input into an improved long sequence prediction network. This network captures local temporal dependencies by embedding a bidirectional long short-term memory network, filters globally associated features using a sparse attention mechanism, achieves nonlinear fusion using a gating mechanism, and compresses the sequence length through a convolution-pooling cascade structure to balance accuracy and efficiency.

[0007] Finally, based on a hard-shared network architecture, TECNFormer, composed of a temporal enhanced convolutional module and an improved long sequence prediction network, is used as the shared feature extraction layer. Multiple task branches for predicting electricity, cooling, and heating loads are set up. Through joint optimization of multi-objective loss functions with equal weights, collaborative prediction across energy loads is achieved, effectively suppressing error propagation and improving prediction robustness.

[0008] This method is not only applicable to multi-load forecasting of integrated energy systems, but can also be extended to other non-stationary time-series scenarios, providing high-precision decision support for applications such as energy dispatch and energy storage optimization.

[0009] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0010] A comprehensive energy system multi-element load forecasting method based on two-level decomposition and reconstruction and TECNFormer includes:

[0011] The standardized input data obtained after preprocessing of historical multi-energy load sequences of integrated energy systems, including electricity, cooling, and heating load sequences, as well as meteorological factors and temporal characteristics;

[0012] A two-level mode decomposition is performed on the standardized input data: the first level obtains intrinsic mode components and residuals through robust empirical mode decomposition, and reconstructs an initial candidate component set; the second level performs continuous variational mode decomposition on the initial candidate component set to obtain a fine-grained candidate component set; based on a multi-index fusion mechanism of frequency information, correlation degree and complexity, the fine-grained candidate component set is adaptively classified into dynamic components of different categories;

[0013] The dynamic components are input into the temporal enhancement convolution module, and multi-scale local features are extracted by combining various heterogeneous differential convolution operators with causal convolution.

[0014] Using the TECNFormer, which consists of the temporal enhancement convolutional module and the improved long sequence prediction network, as a unified shared feature extraction layer, the multi-scale local features, meteorological factors, and temporal features are processed by adaptive position encoding with learnable periodic parameters and then input into the TECNFormer. The improved long sequence prediction network integrates a bidirectional long short-term memory network and a sparse attention mechanism to capture long-range dependencies and local details, thereby obtaining joint modeling features.

[0015] Based on the hard shared network architecture, a multi-task branch is set at the output end of the shared feature extraction layer of the TECNFormer to synchronously output the prediction results of electricity, cooling and heating loads.

[0016] Furthermore, the preprocessing includes: identifying outliers using a sliding window outlier detection method based on periodic features, filling out the outliers with the mean of neighboring data; performing normalization on the filled sequence to eliminate dimensional differences; and using an analysis method integrating Spearman correlation coefficient, Pearson correlation coefficient, and Kendall correlation coefficient to calculate the average of the absolute values ​​of the three types of correlation coefficients as a comprehensive correlation index, screening features that meet a preset threshold for the comprehensive correlation index, and obtaining the standardized input data.

[0017] Furthermore, the two-layer mode decomposition includes:

[0018] In the first layer of robust empirical mode decomposition, the standardized input data is first mirror-symmetric extended, and the sampling points of the preset proportion are reversed and arranged to the two ends of the sequence to form the extended signal. Then, the extended signal is iteratively screened, and the upper and lower envelopes are fitted by cubic spline interpolation and the intermediate signal is updated. When the preset objective function condition is met for several consecutive iterations, the extraction of intrinsic mode components is terminated. The effective points are truncated and the residuals are updated until the residuals meet the monotonicity or energy threshold. The intrinsic mode components and residuals are obtained and reconstructed into the initial candidate component set.

[0019] In the second-level continuous variational mode decomposition, the number of modes, penalty factor and Lagrange multiplier are first set, and the solution is obtained by alternating iteration in the Fourier transform domain. When the convergence error is less than the preset threshold, the process stops. For candidate components whose frequency range after decomposition does not meet the preset frequency separation accuracy, the number of modes is reset and variational mode decomposition is performed again until all candidate components meet the preset frequency separation accuracy, thus obtaining the fine-grained candidate component set.

[0020] Furthermore, the multi-indicator fusion mechanism includes:

[0021] Frequency information is constructed by Hilbert transform to create an analytical signal and extracts instantaneous phase for quantization. The correlation degree is obtained by weighted fusion of self-correlation degree and cross-component nonlinear correlation degree. The self-correlation degree is calculated based on the average value of a lag scale set covering short-term to intraday time intervals, and the nonlinear correlation degree is calculated by using the maximum information coefficient to calculate the average correlation between candidate components. When weighting, the self-correlation degree is given a higher weight to highlight the time period characteristics. The complexity is quantified by multi-scale permutation entropy to quantify the temporal nonlinear characteristics.

[0022] During classification, candidate components that meet the high threshold of any two of the frequency information, correlation degree, and complexity indicators are classified as high dynamic components, those that meet the low threshold of any two indicators are classified as low dynamic components, and the rest are classified as medium dynamic components.

[0023] Furthermore, the temporal enhancement convolution module employs five types of heterogeneous differential convolution operators: central differential convolution, rotational differential convolution, edge differential convolution, horizontal differential convolution, and vertical differential convolution. These are combined with standard convolution to form a six-type convolution operator combination. By stacking temporally dilated convolution blocks with a preset exponential dilation rate of 2, learnable weights are used to fuse the outputs of the six types of convolution operators. Simultaneously, residual connections are used to alleviate the gradient vanishing problem, and causal convolution is used to avoid future information leakage, thereby achieving multi-scale local feature extraction.

[0024] Furthermore, the improved long sequence prediction network specifically comprises:

[0025] Adaptive positional encoding introduces learnable periodic parameters and dynamically adjustable decay rates, and combines temporal encoding to fuse and encode multi-scale local features and exogenous features; a bidirectional long short-term memory network is embedded to capture local temporal information from the past to the present and from the future to the present, and a sparse attention mechanism is used to filter high-information features to capture global dependencies. The bidirectional temporal information and global dependency features are nonlinearly fused through a gating mechanism.

[0026] The encoder employs a convolution-pooling cascade structure, compressing the sequence length layer by layer through activation functions and max pooling; the decoder uses masked multi-head self-attention and cross-attention, combined with future information masks to ensure that predictions rely only on historical data, achieving joint long-term and short-term modeling.

[0027] Furthermore, for the multi-task branches of electricity, cooling, and heating load prediction based on the hard shared network architecture, in order to coordinate the prediction accuracy of each task and optimize the overall performance of the model, a total loss function of multi-task learning is adopted.

[0028] The total loss function is the average of the loss values ​​of the three load prediction tasks: electricity, cooling, and heating, with equal loss weights for each task. By jointly minimizing the total loss function, the network parameters of the shared feature extraction layer of TECNFormer and the corresponding branches of each load prediction task are adjusted synchronously to ensure the consistency and accuracy of the multi-task output.

[0029] And, a multi-dimensional load forecasting system for integrated energy systems based on two-layer decomposition and reconstruction and TECNFormer, comprising:

[0030] The data acquisition and preprocessing module is used to acquire historical multi-energy load sequences, meteorological factors, and temporal characteristics of the integrated energy system, including electricity, cooling, and heating load sequences, and to perform preprocessing on the historical multi-energy load sequences to output standardized input data.

[0031] The two-layer decomposition and classification module is used to perform two-layer mode decomposition on the standardized input data: the first layer obtains intrinsic mode components and residuals through robust empirical mode decomposition, and reconstructs an initial candidate component set; the second layer performs continuous variational mode decomposition on the initial candidate component set to obtain a fine-grained candidate component set; based on a multi-index fusion mechanism of frequency information, correlation degree and complexity, the fine-grained candidate component set is adaptively classified into dynamic components of different categories;

[0032] The temporal enhancement module is used to input the dynamic components into the temporal enhancement convolution module respectively, and extract multi-scale local features by combining multiple heterogeneous differential convolution operators with causal convolution.

[0033] The shared feature extraction layer, composed of the temporal enhancement convolutional module and the improved long sequence prediction network, inputs multi-scale local features, meteorological factors and temporal features into TECNFormer after adaptive position encoding with learnable periodic parameters. The improved long sequence prediction network integrates a bidirectional long short-term memory network and a sparse attention mechanism to capture long-range dependencies and local details, thereby obtaining joint modeling features.

[0034] The hard shared network architecture sets up a multi-task branch at the output of the shared feature extraction layer of the TECNFormer to synchronously output the prediction results of electricity, cooling and heating loads.

[0035] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0036] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0037] Compared with the prior art, the present invention and its preferred embodiments have at least the following beneficial effects:

[0038] By combining robust empirical mode decomposition with continuous variational mode decomposition into a two-layer decomposition and reconstruction framework, along with a component classification mechanism that integrates multiple indicators such as frequency, correlation, and complexity, the model effectively alleviates the mode aliasing and parameter sensitivity problems of traditional decomposition methods. This makes the decoupling of the load sequence more accurate, the physical meaning of each component clearer, and improves the interpretability and engineering reliability of the prediction model.

[0039] The Temporal Enhanced Convolutional Module (TECN) integrates multiple heterogeneous differential convolution operators with temporal dilation, causal convolution, and residual connections, enhancing the ability to extract multimodal features such as mutations, periods, and trends in the load sequence. It suppresses spurious correlations induced by non-stationary noise from the source, significantly improving the model's noise resistance and mutation capture ability.

[0040] The improved long sequence prediction network, through the collaborative design of adaptive spatiotemporal location encoding, bidirectional temporal modeling units and sparse attention mechanisms, combined with a convolution-pooling cascade structure, can efficiently capture long-range dependencies while taking into account local detail characterization and computational efficiency, thus solving the problem of balancing accuracy and speed under long prediction windows.

[0041] The hard-shared multi-task learning architecture makes TECNFormer a unified shared feature extraction layer, which fully explores the time-varying correlations among multiple energy loads such as electricity, cooling, and heating. It avoids information fragmentation and spurious correlations caused by independent component modeling, effectively suppresses error propagation in the decomposition and modeling process, and improves the overall accuracy of multi-load joint prediction.

[0042] The collaborative design of the overall solution gives it strong robustness and generalization ability, making it more robust to situations such as missing disturbances in the load sequence and seasonal weather distribution drift. In addition to comprehensive energy load forecasting, it can also be adapted to other non-stationary time series forecasting scenarios, and has good versatility and scalability. It can provide reliable support for various engineering applications such as economic dispatch, energy storage coordination, and energy use optimization of integrated energy systems. Attached Figure Description

[0043] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0044] Figure 1 This is a diagram of the TECNFormer network architecture according to an embodiment of the present invention;

[0045] Figure 2This is a diagram illustrating the two-layer decomposition and reconstruction architecture of an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the improved Informer principle according to an embodiment of the present invention;

[0047] Figure 4 This is a diagram of a multi-task learning framework based on hard sharing, as described in an embodiment of the present invention. Detailed Implementation

[0048] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0049] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0050] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0051] This invention addresses the problems in existing load forecasting, such as strong non-stationarity, mode aliasing, difficulty in balancing efficiency with long-range dependence, insufficient noise robustness, and inadequate utilization of cross-component and cross-energy correlations. First, it constructs a two-layer decomposition and reconstruction framework based on robust empirical mode decomposition and continuous variational mode decomposition. It then uses multiple indicators—frequency, correlation degree, and complexity—to adaptively classify components into high, medium, and low dynamic ranges, significantly alleviating the mode aliasing and parameter sensitivity issues of traditional single-stage decomposition and improving the physical interpretability and decoupling accuracy of the components. Second, it develops a novel deep learning neural network, TECNForemer, and designs a temporal enhanced convolutional network containing five classes of heterogeneous differential convolution operators. The TECN (Technical Network) module combines multi-scale fusion and causal convolution to enhance the extraction of multimodal features such as mutations, periods, and trends, suppressing spurious correlations induced by non-stationary noise at the source and improving the input quality of long-range modeling. Furthermore, adaptive spatiotemporal location encoding and bidirectional temporal modeling are introduced into the improved Informer, working in conjunction with sparse attention to simultaneously capture long-range dependencies and local details, overcoming the shortcomings of fixed-location encoding's poor adaptability to irregular periods and the susceptibility of pure global attention to noise and computational overhead. Finally, hard-shared multi-task learning is employed, with task-based output on the TECNFormer shared network combining TECN and the improved Informer. This explicitly characterizes the coupling and information sharing between multiple energy loads and multiple components, avoiding information fragmentation and error propagation caused by independent component modeling, thus achieving a better balance between accuracy, robustness, and computational efficiency.

[0052] Therefore, this invention proposes a load forecasting method based on multi-component decoupling and joint long- and short-term modeling, enabling accurate forecasting of strongly non-stationary, multi-scale coupled integrated energy loads (electricity / cooling / heating, etc.). The specific implementation steps include:

[0053] 1. Data Acquisition and Preprocessing

[0054] 1.1 Obtain historical multi-energy load sequences (electricity, cooling and heating loads), meteorological factors (temperature, humidity, wind speed, irradiance, etc.), holiday and time characteristics (hour, weekday, holiday labels), and optional service-side characteristics (electricity price, equipment status, etc.).

[0055] 1.2 Anomaly detection is performed using the sliding window Z-score method based on periodic features, and the mean of neighboring data is used for filling. Furthermore, the dimensional differences are eliminated through max-min normalization.

[0056] 1.3 Feature correlation screening is performed using an analysis method that integrates Spearman, Pearson and Kendall correlation coefficients. The average of the absolute values ​​of the three is used as the comprehensive correlation index. Features with a comprehensive correlation index greater than 0.3 are selected to reduce input redundancy and noise coupling.

[0057] 2 Candidate component classification

[0058] Candidate components are classified based on frequency information, correlation, and complex quantitative characteristics through a multi-index fusion mechanism. Candidate components that meet the high threshold of any two categories of indices are classified as high-dynamic components, representing high-frequency fluctuations dominated by random disturbances or inter-energy flow interactions; candidate components that meet the low threshold of any two categories of indices are classified as low-dynamic components, corresponding to slowly changing load base load trends; candidate components that do not meet the high or low dynamic conditions are classified as medium-dynamic components, possessing a certain degree of periodicity and complexity, reflecting medium-frequency periodic fluctuation characteristics.

[0059] Frequency information is analyzed using adaptive time-frequency analysis employing the Hilbert Transform (HT). This involves constructing an analytic signal, extracting the instantaneous phase, and calculating the instantaneous frequency mean f. ’ (t) Frequency information of the quantized component:

[0060]

[0061] Where H[∙] represents the Hilbert transform operation.

[0062] The correlation is characterized from two aspects: autocorrelation and cross-component nonlinear correlation. Autocorrelation is expressed as the average value of the autocorrelation coefficient (AC) across multiple lag scales. avg To quantify the lag correlation of candidate components, the lag scale set is set as τ={1,4,12,24,48,96} (corresponding to time intervals from 15 minutes to 24 hours):

[0063]

[0064] Among them, DC ` The mean of the candidate components.

[0065] Nonlinear correlation is quantified by the maximum information coefficient (MIC) as the mean MIC value between the current candidate component and the remaining candidate components. avg The normalized mutual information of the joint probability distribution is calculated using the grid partitioning method:

[0066]

[0067] Among them, G x and G y Divide the two-dimensional network into sample sizes; B(N) = [n 1 / 2 ] represents the upper limit of the number of networks; I(∙) represents mutual information.

[0068] The formula for calculating the correlation coefficient (Corr) is as follows, used to integrate autocorrelation and nonlinear correlation:

[0069]

[0070] Where, ω AC= 0.7 and ω MIC =0.3 represents the weights of the AC and MIC indicators, respectively, prioritizing time cycle characteristics; MIC 1,avg MIC represents the mean MIC among candidate components of the same type of load; 2,avg and MIC 3,avg This represents the mean MIC among candidate components of cross-class loads.

[0071] The complexity is quantified using Multiscale Permutation Entropy (MPE) to determine the temporal nonlinear complexity. Coarse-grained sequences of candidate components at different time scales τ are constructed, and these sequences are arranged in ascending order of continuous d... MPE A permutation pattern δ is generated from each point, and the mean permutation entropy MPE is calculated based on the probability distribution p(π). avg :

[0072] .

[0073] 3. Two-layer decomposition and reconstruction framework

[0074] The two-layer decomposition and reconstruction architecture provided in this embodiment is as follows: Figure 2 As shown, it specifically includes:

[0075] 3.1 Perform the first-level robust empirical mode decomposition on the original load sequence to obtain several intrinsic mode functions (IMFs) and residuals (RES).

[0076] Step 1. Signal extension.

[0077] The original time series is mirror-symmetrically extended, and the first M=0.1N sampling points are reversed and placed at both ends of the sequence to form an extended signal x of length N+2M. extended (t).

[0078] Step 2. Iterative sieving.

[0079] Initialize RESr0(t) = x extended (t) and IMF index k=1, when RES is non-monotonic and energy is above the threshold, the upper and lower envelopes e are fitted by cubic spline interpolation. up (t) and e low (t) and update the intermediate signal h k+1 (t)=h k (t)-(e up (t)+e low (t)) / 2.

[0080] Define the objective function f k (t)=RMS(h k (t)+|EK(h k (t))|), when f is satisfied in three consecutive iterations k-2 (t) <f k-1 (t) <f k When (t), terminate the current IMF extraction and record the IMF. k (t)=h k-2 (t).

[0081] Subtract the current IMF from the extended time series and extract the N valid points in the middle to update RESr. k (t) until it is a monotonic sequence or the RMS is less than 0.1% of the original signal.

[0082] 3.2 Reconstruct the first-level candidate components into a candidate component set.

[0083] 3.3 For the candidate component set, a second-level continuous variational mode decomposition is performed to obtain finer-grained candidate components. The initial number of modes, the penalty factor controlling the bandwidth constraint strength, and the Lagrange multipliers are set, and the solution is obtained iteratively in the Fourier transform domain. When the convergence error is less than a threshold... Stop at this point. For candidate components with a wide frequency range, reset the mode number and decompose sequentially until the preset frequency separation accuracy is met.

[0084]

[0085] Among them, u k (t) represents the k-th candidate component; w k The center frequency is K; the number of candidate components is K; and λ is the Lagrange multiplier. Frequency domain partitioning is achieved by minimizing the sum of the generalized bandwidths of the candidate components.

[0086] 3.4 Reconstruct the second-layer candidate components into a candidate component set.

[0087] 4. Temporal Augmentation Convolution Module

[0088] 4.1 The TECNFormer network (i.e., time-series augmented convolutional module) architecture provided in this embodiment is as follows: Figure 1 As shown, each candidate component is input into TECN, which includes five families of heterogeneous differential convolution operators, denoted as Central Difference Convolution (CDC), Angular Difference Convolution (ADC), Radial Difference Convolution (RDC), and Horizontal / Vertical Difference Convolution (HDC / VDC). These operators, combined with standard convolution, preserve the original structural information, enabling differentiated extraction of multimodal features.

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095] Among them, w CDC w ADC w RDC w HDC and w VDC For the convolution kernels of CDC, ADC, RDC, HDC, and VDC, w s ∈ It is a two-dimensional convolution kernel, with the kernel center at (i c i c );k c For kernel size; σ ADC σ is the angle modulation parameter; RDC,1 and σ RDC,2 For rotation modulation parameters; δ w σ is the directional modulation function; HDC and σ VDC These are the modulation parameters for the horizontal and vertical directions.

[0096] 4.2 Temporal Enhancement Blocks with an exponential inflation rate of 2 are stacked, and learnable weights α are used to... c It integrates the outputs of six types of convolution operators, uses causal convolution to avoid future information leakage, and combines residual connections to alleviate gradient vanishing.

[0097]

[0098] .

[0099] 5. Long-term and short-term joint modeling and cross-component fusion

[0100] The improved Informer principle provided in this embodiment is as follows: Figure 3 As shown, specifically:

[0101] 5.1 Based on the standard sparse attention of the Informer encoder, a learnable periodic parameter T is designed. PE Dynamically adjust the attenuation rate τ PE and periodic modulation factor μ PE Furthermore, it integrates time coding to achieve explicit modeling of multi-scale periodic features:

[0102]

[0103]

[0104] Where PE is the location code; TE is the time code; p is the location index; d is the embedding dimension; t is the position ... position index; t is the position index; d is the embedding dimension; t is the position index; t is the position index; d is the position index; t is the position index; t is the position index; d is the embedding dimension; t is the position index; t is the feature For time stamps.

[0105] 5.2 Temporal enhancement features, meteorological features, and temporal features are encoded in spatiotemporal location and then input into Informer. The sparse attention mechanism and bidirectional temporal dependency are used to filter cross-time period associations.

[0106]

[0107] in, This is the embedding vector.

[0108] 5.3 The sparse attention mechanism reduces computational complexity by probabilistically filtering high-information query vectors. To enhance the bidirectional modeling capability of local temporal dependencies, a bidirectional long short-term memory network (Bi-LSTM) is embedded in the encoder, using forward hidden states ( ) and backward hidden state ( This captures information from the past to the present and from the future to the present, respectively. The bidirectional features and the globally correlated features output by the sparse attention mechanism are non-linearly fused through a gating mechanism.

[0109]

[0110]

[0111]

[0112]

[0113] Where Attn represents the sparse attention mechanism; M is the sparsity measure; Q, K, and V are the query, key, and value matrices, respectively; and L is the length of the input sequence. The filtered query vector; w g and b g denoted as the gating weight matrix and bias vector, respectively; g is the gating adjustment ratio; and f is the fused feature.

[0114] 5.5 The convolution-pooling cascade structure compresses the sequence length layer by layer, reducing computational complexity while avoiding information dilution.

[0115]

[0116] Among them, X Encoder,e represents the output feature of the e-th encoder layer.

[0117] 5.6 The decoder uses masked multi-head self-attention Attn. self Capture local information from the decoder input and cross-attention Attn corss By fusing global features from the encoder and using future information masks to ensure that predictions rely solely on historical data, causal time-series predictions are achieved.

[0118]

[0119] 6. Multi-task learning output

[0120] This embodiment adopts Figure 4 The illustrated multi-task learning framework based on hard sharing uses TECNFormer as the shared network. At the output, multiple task branches are set to predict the reconstructed components of electrical, cooling, and heating loads respectively. The loss function adopts a weighted multi-objective form, with each task's loss function weight fixed at 1. The total loss function is Loss. total Defined as the loss of P tasks.p The sum of:

[0121] .

[0122] 7. Evaluation of Error Indicators

[0123] This embodiment uses the Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) to comprehensively analyze the model performance.

[0124]

[0125] .

[0126] Based on the solution provided above in this embodiment, its significant effects include:

[0127] 1. Robust decoupling and improved interpretability: The system adopts a two-level decomposition of REMD+SVMD and adaptively divides the components into HDC, MDC and LDC based on multiple indicators such as "frequency-relevance-complexity". This significantly alleviates modal mixing and parameter sensitivity, and makes the physical meaning of the components clearer, which is easier for the operation and maintenance side to understand and review.

[0128] 2. Long-term and short-term collaboration, balancing accuracy and efficiency: TECN extracts multimodal features such as local mutations, cycles, and trends; the improved Informer introduces adaptive spatiotemporal location encoding and Bi-LSTM in collaboration with sparse attention to achieve simultaneous characterization of long-term dependencies and local details, reducing inference complexity and latency while ensuring accuracy.

[0129] 3. Error propagation suppression: During the decomposition, classification and reconstruction process, cross-component information sharing and structured constraints are used to reduce the cascading amplification of parameters / noise in the decomposition stage on subsequent prediction and reconstruction.

[0130] 4. Cross-component / cross-energy coupling characterization: Hard shared multi-task learning outputs tasks on the shared representation, making full use of multi-energy loads such as electricity, cooling and heating and the time-varying correlation between different components, avoiding information fragmentation and spurious correlation caused by independent modeling.

[0131] 5. Strong robustness and generalization ability: Heterogeneous differential convolution (CDC, ADC, RDC, HDC and VDC) together with causal convolution and multi-scale fusion enhance the noise resistance and mutation capture ability, making it more robust to missing perturbations and seasonal weather distribution drift, and more friendly to cross-scene transfer.

[0132] 6. General and Scalable: In addition to comprehensive energy load, it is also applicable to other non-stationary time series (such as renewable power output, industrial processes, traffic flow, etc.), and supports multi-step forecasting, rolling forecasting and parallel computing expansion.

[0133] The solutions provided by this invention can be used for forecasting and rolling updates of diverse loads such as electricity, cooling, and heating in ultra-short-term, short-term, day-ahead, and intraday periods; provide high-precision priors for the economic dispatch, energy coupling, and energy storage synergy of energy systems (IES) in industrial parks, urban energy stations, data centers, hospitals, and airports; support the tracking of wind and solar power output, and the start-up / ramp-up and standby arrangements of pumped storage and conventional hydropower, improving the absorption rate and system flexibility; provide forecasting baselines and execution references for load aggregation, peak shaving and valley filling, interruptible loads, and price arbitrage strategies; support energy consumption planning, peak-valley electricity price optimization, energy storage charging and discharging strategies, and contract capacity assessment for building and industrial park energy management systems; based on component and residual anomaly identification, it can be used for equipment health diagnosis, energy consumption anomaly alarms, and fault location; it can be used for capacity expansion assessment, resilience analysis, and carbon emission assessment of substations / energy storage / cooling stations; and it can provide forecasting support and uncertainty boundaries for day-ahead / real-time market quotations, standby, and frequency regulation.

[0134] Finally, here is a test example implementation:

[0135] This invention can be implemented using the PyTorch deep learning framework, with an NVIDIA RTX 4090 GPU (24GB VRAM), an Intel Core i9-13900K CPU, and 128GB of RAM. The dataset is divided into a training set (80%) and a test set (20%) in chronological order. A 5-fold hierarchical cross-validation method is used to evaluate the model's generalization ability. For each validation, four subsets are randomly selected from the training set for training, and the remaining subset is used for validation, ensuring a balanced sample distribution. A sliding window is used to construct input and output sample pairs for single-step prediction. To avoid future information leakage, the validation process strictly follows the chronological order, ensuring that all test set samples are within the time frame of the training set. Model hyperparameters are determined iteratively through grid search combined with validation set performance, using the Adam optimizer. A hard-shared MTL architecture is adopted, with shared bottom-level feature extraction layers and independently configured fully connected output layers for each load component. The loss function is the sum of the root square error (RSE) of each task, with weights evenly distributed.

[0136] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0137] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0138] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0140] This invention is not limited to the preferred embodiment described above. Anyone inspired by this invention can derive various other forms of a multi-dimensional load forecasting method for integrated energy systems based on two-layer decomposition and reconstruction and TECNFormer. All equivalent variations and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. A multi-element load forecasting method for integrated energy systems based on two-layer decomposition and reconstruction and TECNFormer, characterized in that, include: The standardized input data obtained after preprocessing of historical multi-energy load sequences of integrated energy systems, including electricity, cooling, and heating load sequences, as well as meteorological factors and temporal characteristics; A two-level mode decomposition is performed on the standardized input data: the first level obtains intrinsic mode components and residuals through robust empirical mode decomposition, and reconstructs an initial candidate component set; the second level performs continuous variational mode decomposition on the initial candidate component set to obtain a fine-grained candidate component set; based on a multi-index fusion mechanism of frequency information, correlation degree and complexity, the fine-grained candidate component set is adaptively classified into dynamic components of different categories; The dynamic components are input into the temporal enhancement convolution module, and multi-scale local features are extracted by combining various heterogeneous differential convolution operators with causal convolution. Using the TECNFormer, which consists of the temporal enhancement convolutional module and the improved long sequence prediction network, as a unified shared feature extraction layer, the multi-scale local features, meteorological factors, and temporal features are processed by adaptive position encoding with learnable periodic parameters and then input into the TECNFormer. The improved long sequence prediction network integrates a bidirectional long short-term memory network and a sparse attention mechanism to capture long-range dependencies and local details, thereby obtaining joint modeling features. Based on the hard shared network architecture, a multi-task branch is set at the output end of the shared feature extraction layer of the TECNFormer to synchronously output the prediction results of electricity, cooling and heating loads.

2. The integrated energy system multi-element load forecasting method based on two-layer decomposition and reconstruction and TECNFormer as described in claim 1, characterized in that: The preprocessing includes: identifying outliers using a sliding window outlier detection method based on periodic features; filling out the outliers using the mean of neighboring data; performing normalization on the filled sequence to eliminate dimensional differences; and using an integrated analysis method combining Spearman correlation coefficient, Pearson correlation coefficient, and Kendall correlation coefficient to calculate the average of the absolute values ​​of the three types of correlation coefficients as a comprehensive correlation index, screening features that meet a preset threshold for the comprehensive correlation index, and obtaining the standardized input data.

3. The integrated energy system multi-element load forecasting method based on two-layer decomposition and reconstruction and TECNFormer as described in claim 1, characterized in that: The bi-level mode decomposition includes: In the first layer of robust empirical mode decomposition, the standardized input data is first mirror-symmetric extended, and the sampling points of the preset proportion are reversed and arranged to the two ends of the sequence to form the extended signal. Then, the extended signal is iteratively screened, and the upper and lower envelopes are fitted by cubic spline interpolation and the intermediate signal is updated. When the preset objective function condition is met for several consecutive iterations, the extraction of intrinsic mode components is terminated. The effective points are truncated and the residuals are updated until the residuals meet the monotonicity or energy threshold. The intrinsic mode components and residuals are obtained and reconstructed into the initial candidate component set. In the second-level continuous variational mode decomposition, the number of modes, penalty factor and Lagrange multiplier are first set, and the solution is obtained by alternating iteration in the Fourier transform domain. When the convergence error is less than the preset threshold, the process stops. For candidate components whose frequency range after decomposition does not meet the preset frequency separation accuracy, the number of modes is reset and variational mode decomposition is performed again until all candidate components meet the preset frequency separation accuracy, thus obtaining the fine-grained candidate component set.

4. The integrated energy system multi-element load forecasting method based on two-layer decomposition and reconstruction and TECNFormer as described in claim 1, characterized in that: The multi-indicator fusion mechanism includes: Frequency information is constructed by Hilbert transform to create an analytical signal and extracts instantaneous phase for quantization. The correlation degree is obtained by weighted fusion of self-correlation degree and cross-component nonlinear correlation degree. The self-correlation degree is calculated based on the average value of a lag scale set covering short-term to intraday time intervals, and the nonlinear correlation degree is calculated by using the maximum information coefficient to calculate the average correlation between candidate components. When weighting, the self-correlation degree is given a higher weight to highlight the time period characteristics. The complexity is quantified by multi-scale permutation entropy to quantify the temporal nonlinear characteristics. During classification, candidate components that meet the high threshold of any two of the frequency information, correlation degree, and complexity indicators are classified as high dynamic components, those that meet the low threshold of any two indicators are classified as low dynamic components, and the rest are classified as medium dynamic components.

5. The multi-element load forecasting method for integrated energy systems based on two-layer decomposition and reconstruction and TECNFormer as described in claim 1, characterized in that: The temporal enhancement convolution module employs five heterogeneous differential convolution operators: central differential convolution, rotational differential convolution, edge differential convolution, horizontal differential convolution, and vertical differential convolution. These are combined with standard convolution to form a six-operator combination. By stacking temporally dilated convolution blocks with a preset exponential dilation rate of 2, learnable weights are used to fuse the outputs of the six convolution operators. Simultaneously, residual connections are used to alleviate the gradient vanishing problem, and causal convolution is used to avoid future information leakage, thereby achieving multi-scale local feature extraction.

6. The multi-element load forecasting method for integrated energy systems based on two-layer decomposition and reconstruction and TECNFormer as described in claim 1, characterized in that: The improved long sequence prediction network is specifically as follows: Adaptive positional encoding introduces learnable periodic parameters and dynamically adjustable decay rates, and combines temporal encoding to fuse and encode multi-scale local features and exogenous features; a bidirectional long short-term memory network is embedded to capture local temporal information from the past to the present and from the future to the present, and a sparse attention mechanism is used to filter high-information features to capture global dependencies. The bidirectional temporal information and global dependency features are nonlinearly fused through a gating mechanism. The encoder employs a convolution-pooling cascade structure, compressing the sequence length layer by layer through activation functions and max pooling; The decoder employs masked multi-head self-attention and cross-attention, combined with future information masks to ensure that predictions rely solely on historical data, thus achieving joint long-term and short-term modeling.

7. The multi-element load forecasting method for integrated energy systems based on two-layer decomposition and reconstruction and TECNFormer as described in claim 1, characterized in that: For the multi-task branch of electricity, cooling and heating load prediction based on the hard shared network architecture, in order to coordinate the prediction accuracy of each task and optimize the overall performance of the model, the total loss function of multi-task learning is adopted. The total loss function is the average of the loss values ​​of the three load prediction tasks: electricity, cooling, and heating, with equal loss weights for each task. By jointly minimizing the total loss function, the network parameters of the shared feature extraction layer of TECNFormer and the corresponding branches of each load prediction task are adjusted synchronously to ensure the consistency and accuracy of the multi-task output.

8. A multi-element load forecasting system for integrated energy systems based on two-layer decomposition and reconstruction and TECNFormer, characterized in that, include: The data acquisition and preprocessing module is used to acquire historical multi-energy load sequences, meteorological factors, and temporal characteristics of the integrated energy system, including electricity, cooling, and heating load sequences, and to perform preprocessing on the historical multi-energy load sequences to output standardized input data. The two-layer decomposition and classification module is used to perform two-layer mode decomposition on the standardized input data: the first layer obtains intrinsic mode components and residuals through robust empirical mode decomposition, and reconstructs an initial candidate component set; the second layer performs continuous variational mode decomposition on the initial candidate component set to obtain a fine-grained candidate component set; based on a multi-index fusion mechanism of frequency information, correlation degree and complexity, the fine-grained candidate component set is adaptively classified into dynamic components of different categories; The temporal enhancement module is used to input the dynamic components into the temporal enhancement convolution module respectively, and extract multi-scale local features by combining multiple heterogeneous differential convolution operators with causal convolution. The shared feature extraction layer, composed of the temporal enhancement convolutional module and the improved long sequence prediction network, inputs multi-scale local features, meteorological factors and temporal features into TECNFormer after adaptive position encoding with learnable periodic parameters. The improved long sequence prediction network integrates a bidirectional long short-term memory network and a sparse attention mechanism to capture long-range dependencies and local details, thereby obtaining joint modeling features. The hard shared network architecture sets up a multi-task branch at the output of the shared feature extraction layer of the TECNFormer to synchronously output the prediction results of electricity, cooling and heating loads.

9. A computer device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1-7.

Images