Method and system for multi-energy load forecasting in the absence of historical data for an integrated energy system

Abstract

A multi-energy load forecasting method, a multi-energy load forecasting system, an electronic device, a program, and a storage medium are provided that realize accurate long-term forecasting of multi-energy loads in a target integrated energy system under conditions where no historical load data is available. [Solution] A multi-energy load forecasting method for an integrated energy system without historical data involves obtaining the meteorological characteristics of a target complex and the cooling, heating, electricity, and gas historical data of a source domain group complex, preprocessing the obtained data, performing cross-correlation and generalization ability analysis of the complex on the preprocessed cooling, heating, electricity, and gas historical data of the source domain group complex, determining appropriate source domain data, constructing a multi-energy load forecasting model, training the model based on the source domain data according to the Metas training policy, obtaining a trained forecasting model, and inputting the preprocessed meteorological characteristics of the target complex into the forecasting model to obtain a forecast result.

Description

[Technical Field]

[0001] (CROSS-REFERENCE TO RELATED APPLICATIONS) This application claims priority to a Chinese patent application bearing application number 202510412447.8 and entitled "Method and system for multi-energy load forecasting without historical data for integrated energy system," filed with the State Intellectual Property Office of the People's Republic of China on April 3, 2025, the entire contents of which are incorporated by reference into and constitute a part of the present invention for all purposes.

[0002] The present invention relates to the technical field of load forecasting, and more particularly to a method and system for multi-energy load forecasting in the absence of historical data for an integrated energy system. [Background technology]

[0003] The statements in this section are merely intended to provide background information related to the present invention and do not necessarily constitute prior art.

[0004] With the rapid development of integrated energy systems (IES), accurate prediction of multi-energy loads has become an important requirement for optimizing system operation. However, in practical applications, acquiring comprehensive multi-energy load history data generally requires a lot of time and capital. Furthermore, due to data privacy and other reasons, some information cannot be made public, and the size of the available data set is limited. Traditional deep learning methods face serious data dependency bottlenecks, posing a huge challenge to IES load forecasting. In response, many scholars both in China and abroad have adopted transfer learning methods to address this challenge.

[0005] Performing correlation analysis on load sequences to aid in the selection of an appropriate source domain is one of the most important steps before transfer learning. Some scholars consider the complex linear and nonlinear characteristics of load data, thereby improving the correlation analysis and further reducing the amount of target domain data required. However, in extreme situations where the community is not fully developed, the target system may not have any historical load data. Traditional source domain selection relies on target domain correlation analysis, but feature comparison is completely ineffective under data-free conditions. Some scholars have proposed transfer learning methods based on generative adversarial networks (GANs), fine-tuning policies, and meta-learning frameworks, but they still require a small amount of target domain data and have obvious shortcomings in extracting complex time-series features and co-optimizing multiple source domains.

[0006] For these reasons, scholars in China and abroad have proposed various solutions to achieve high-precision multi-energy load forecasting under conditions where load data is extremely scarce. However, when there is no historical load data for newly constructed parks, traditional transfer learning techniques face a double bottleneck.

[0007] (1) It is difficult to select the source domain. In transfer learning, the similarity analysis between the source domain and the target domain strongly depends on the target domain data, so the conventional correlation analysis method is ineffective under the condition of no data.

[0008] (2) Lack of cross-domain generalization ability: Cross-domain parameter transfer does not fully consider the dynamic variations of multi-source domains, resulting in poor parameter transfer efficiency and large prediction errors. Summary of the Invention

[0009] To solve the above problems, the present invention proposes a method and system for multi-energy load forecasting in the absence of historical data for an integrated energy system. Based on the Tnet algorithm and an improved meta-learning policy, a multi-energy load forecasting model based on transfer learning is constructed to achieve accurate long-term forecasting of multi-energy loads in a target integrated energy system in the absence of historical load data.

[0010] To achieve the above objectives, the present invention adopts the following technical solutions.

[0011] In a first aspect, the present invention provides a method for producing a composition comprising: Obtaining weather characteristics of the target complex and cooling, heating, electricity, and gas history data of the source domain group complex, and preprocessing the obtained data; A step of performing cross-correlation and generalization ability analysis of the cooling, heating, electricity, and gas historical data of the source domain group housing complexes after preprocessing to determine appropriate source domain data; Construct a multi-energy load forecasting model, train the model based on source domain data according to the Metas training policy, and adjust the gradient weights based on the validation loss of the inner loop task and the source domain generalization probability to obtain a trained forecasting model; inputting the pre-processed weather characteristics of the target area into a forecasting model to obtain a forecast result; The present invention provides a method for multi-energy load forecasting in the absence of historical data for an integrated energy system, including:

[0012] In an optional embodiment, the Metas training policy specifically: Considering the parameter updates of different source domains, the loss values ​​and weight values ​​of each training task in the inner layer are fully affected by the update function in the outer layer.

[0013] In an alternative embodiment, the multi-energy load prediction model comprises an encoder module and a decoder module.

[0014] In an optional embodiment, the encoder uses a CNN layer to perform feature extraction and dimensionality reduction through convolutional operations, and then sends the compressed data to an LSTM to further capture long-term dependencies in the time series.

[0015] In an optional embodiment, we utilize a multi-head attention mechanism to dynamically weight the relationships at different time steps, thereby giving different hidden state probability weights to the LSTM and focusing attention on the influence of important information related to cooling, heating, electricity, and gas loads.

[0016] In an alternative embodiment, the cluster cross-correlation and generalization ability analysis may specifically include: The Time2vec algorithm is used to convert multiple load time series data for each source domain into two parts: periodic and linear, and then the MIC and MD algorithms are used to process them, respectively. K-means cluster analysis is used to perform generalized potential stratification for the source domains, and low-potential clusters are removed to reduce noise interference. Weighted fusion is performed for MD and MIC, and grid search is used to find the optimal parameter range, thereby obtaining the cross-correlation score of the source domain group and a reasonable source domain range.

[0017] In a second aspect, the present invention provides a method for producing a pharmaceutical composition comprising: a data acquisition and preprocessing module configured to acquire meteorological characteristics of the target complex and historical data of cooling, heating, electricity, and gas of the source domain group complex, and preprocess the acquired data; a source domain data determination module configured to perform a cross-correlation and generalization ability analysis of the cooling, heating, electricity, and gas historical data of the source domain group housing complexes after preprocessing to determine appropriate source domain data; a model building and training module, configured to build a multi-energy load forecasting model, train the model based on source domain data according to a Metas training policy, and adjust gradient weights based on the validation loss and source domain generalization probability of the inner-loop task to obtain a trained forecasting model; a model output module configured to input the preprocessed meteorological characteristics of the target area into a forecasting model to obtain a forecast result; The present invention provides a multi-energy load forecasting system for an integrated energy system without historical data, including:

[0018] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and computer instructions stored in the memory and executable on the processor, the computer instructions, when executed by the processor, completing the method of the first aspect.

[0019] In a fourth aspect, the present invention provides a computer-readable storage medium used to store computer instructions that, when executed by a processor, complete the method of the first aspect.

[0020] In a fifth aspect, the present invention provides a computer program product comprising a computer program that, when executed by a processor, enables completion of the method of the first aspect.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] The method and system for multi-energy load forecasting in the absence of historical data for an integrated energy system of the present invention is based on the Tnet algorithm and an improved meta-learning policy, constructs a multi-head attention optimization encoding-decoding model with transfer learning, achieves prediction breakthrough through three-stage cascade optimization, realizes long-term forecasting of cooling, heating, electricity, and gas loads in the target domain, and significantly improves the accuracy of load forecasting in the absence of data.

[0023] In the method and system for multi-energy load forecasting without historical data for an integrated energy system of the present invention, a source domain selection algorithm (Tnet) is proposed when there is no historical load data for the target domain, which transforms the traditional source domain similarity analysis into a model generalization ability probability evaluation task. Without requiring any target domain historical data, Tnet is used to screen the most appropriate source domain set, realize load forecasting, and overcome the initialization difficulties of transfer learning under data-free constraints.

[0024] The method and system for multi-energy load forecasting without historical data for an integrated energy system of the present invention designs an improved meta-learning training algorithm (Metas), which introduces the loss value and weight information of each inner-loop task during the update of the outer loop of meta-learning, fully utilizes the multi-source domain knowledge provided by Tnet, breaks through the parameter transfer barrier in time series forecasting, and improves the generalization performance of the forecasting model.

[0025] Additional advantages of the invention will be set forth in part in the description that follows, and in part will be obvious from the description, or may be learned by practice of the invention.

[0026] The drawings in the specification that form a part of this invention are intended to provide a further understanding of the invention, and the illustrative embodiments of the invention and their descriptions are intended to interpret the invention and are not intended to unduly limit the invention. [Brief explanation of the drawings]

[0027] [Figure 1] 1 is a flowchart of a multi-energy load forecasting method without history data of an integrated energy system provided in a first embodiment of the present invention; [Figure 2] 1 is a flowchart of the Tnet algorithm provided in the first embodiment of the present invention; [Figure 3] FIG. 1 is a principle diagram of the Metas algorithm provided in the first embodiment of the present invention. [Figure 4] 4A and 4B are framework diagrams of a multi-energy load forecasting model provided in Example 1 of the present invention, in which FIG. 4(a) is an overall framework diagram of the model, and FIG. 4(b) is a coder-decoder framework diagram of the model. DETAILED DESCRIPTION OF THE INVENTION

[0028] The invention will now be further described with reference to the following figures and examples.

[0029] It should be noted that the following detailed description is for illustrative purposes only and is intended to further explain the present invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art.

[0030] It should be noted that the terminology used herein is merely for the purpose of describing specific embodiments and is not intended to limit the exemplary embodiments according to the present invention. For example, unless the context clearly dictates otherwise, the singular forms used herein are intended to include the plural forms, and the terms "comprises," "having," and any variations thereof are intended to cover a non-exclusive inclusion; for example, it should be further understood that a process, method, system, product, or apparatus comprising a series of steps or units is not limited to the steps or units expressly recited, but may also include other steps or units that are not expressly recited or inherent in those processes, methods, products, or apparatus.

[0031] The embodiments and features of the embodiments of the present invention can be combined with each other without any contradiction.

[0032] Example 1 As shown in FIG. 1, this embodiment Obtaining weather characteristics of the target complex and cooling, heating, electricity, and gas history data of the source domain group complex, and preprocessing the obtained data; A step of performing cross-correlation and generalization ability analysis of the cooling, heating, electricity, and gas historical data of the source domain group housing complexes after preprocessing to determine appropriate source domain data; Construct a multi-energy load forecasting model, train the model based on source domain data according to the Metas training policy, and adjust the gradient weights based on the validation loss of the inner loop task and the source domain generalization probability to obtain a trained forecasting model; inputting the pre-processed weather characteristics of the target area into a forecasting model to obtain a forecast result; The present invention provides a method for multi-energy load forecasting in the absence of historical data for an integrated energy system, including:

[0033] To solve the extreme problem of a lack of historical load data for the target domain, this application proposes a novel multi-energy load forecasting method. This method includes a coding-decoding prediction model that combines a new source domain selection algorithm, an improved meta-learning training policy, and a multi-head attention mechanism. It achieves breakthrough prediction through three-stage cascade optimization, thereby achieving accurate long-term forecasts for cooling, heating, electricity, and gas loads in the target integrated energy system. It is applicable to the planning and scheduling of newly constructed complexes or integrated energy systems lacking historical data.

[0034] As shown in Figure 1, the raw data from the source domain group and the target complex first undergo preprocessing, including outlier detection, data normalization, and feature contribution assessment. Next, the Tnet algorithm is used to analyze the cross-correlation and generalization ability of the complexes in the source domain group, transforming the similarity analysis problem into a classification problem and identifying suitable complexes as source domains. Combined with the target complex data preprocessing, an autocorrelation analysis is then performed to determine the input data type and period. Data from multiple complexes is then input into the prediction model, which sequentially extracts features using a convolutional neural network (CNN), a long short-term memory network (LSTM), and a multi-head attention mechanism (MHA). The combined information is then decomposed and extracted through a shared encoding and decoding module, resulting in a joint prediction result. Metas applies the loss and weight values ​​of each task in the inner layer to the outer layer function, further updating the model parameters. Finally, the weather features of the target complex are input into the updated model, enabling multi-energy prediction of cooling, heating, electricity, and gas load data when there is no historical data.

[0035] Here, the Tnet algorithm is used to perform cross-correlation and generalization ability analysis of clusters for a source domain group to find a cluster suitable as a source domain, as follows:

[0036] The Tnet algorithm reconstructs the source domain selection paradigm through a five-stage probabilistic inference framework, and uses five main steps to find suitable source domain data: Time to Vector Embedding (Time2vec), Mahalanobis Distance (MD), Maximal Information Coefficient (MIC), K-means Clustering (K-means), and Bayesian Weighted Probability Averaging Method (BWPA). In Figure 1,

number

number

number

[0037] Figure 2 shows the flowchart. Because the historical cooling, heating, electricity, and gas load data for the target housing complex is unknown, it is crucial to identify a group of similar housing complexes as the source domain. Given the periodicity and linear characteristics of the load data used, the single consideration of linear and nonlinear factors, and the direct use of linear and nonlinear analysis, are inevitably affected by opposing factors. Therefore, the Time2vec algorithm is used to convert the multiple load time series data from each source domain into two parts: periodic and linear, and map the time features into a single high-dimensional space. The periodicity primarily accounts for seasonal changes in load forecasting, while the linear features capture the baseline trend of the load data. This is shown in Equation (1). In this equation, t2v(τ) represents the initial time series, τ represents the time step size, and i represents the embedding dimension. The 0th dimension represents linear features, while the 1st through kth dimensions represent nonlinear features.

[0038]

number

[0039] To perform the joint analysis of linear and non-linear in high-dimensional space, the MD algorithm is used to introduce the covariance matrix, adjust the weights of different features, so that features with different scales and different correlations are fairly handled during the operation. The MIC algorithm is used to divide different grids for different non-linear variables and find the maximum amount of information by different discretization policies. As shown in equations (2) and (3). Furthermore, MD and MIC are used to construct K-means cluster analysis to perform generalization potential stratification on the source domain, and remove low-potential clusters to reduce noise interference. One cluster contains multiple housing estates. If they are used uniformly, the respective generalization capabilities of each housing estate for the target housing estate cannot be considered, and the prediction error is very large. Therefore, weight fusion is performed on MD and MIC according to equation (4), and the optimal parameter range is found using grid search to obtain the mutual correlation score of the source domain group in this cluster and a reasonable source domain range.

[0040]

Number

[0041] Finally, in order to fully utilize the contribution degrees of different source domains to the target domain in the prediction model, weighted mutual information entropy is introduced to construct a Bayesian weighting model, reasonable source domains are used as tags, and a parameter model is constructed by α in equation (4) and substituted into equation (5), and weights are assigned considering the "uncertainty" of each data point. In this way, the conventional probability prediction problem is converted into a classification problem to find the most appropriate housing estate group. Here, X and Y are the data of cooling, heating, electricity, and gas of two different housing estates, both of which are four-dimensional. Σ -1 indicates that it is the inverse matrix of Σ, represents the correlation of each feature between data points, and is a 4×n matrix. G is to divide the data into a×b grids, and I(X,Y|G) is the mutual information estimation value in grid G, and the maximum value is selected from all grids that satisfy a×b < B(n) (usually, B(n) = n 0.6, n is the sample volume), P(M i ) is Model M i is the prior probability of , y is the predicted outcome, and X is the input feature.

[0042] Therefore, in this application, this algorithm can be regarded as a problem of maximizing generalization ability after standardizing data of four different scales: cooling, heating, electricity, and gas. At the same time, the Tnet algorithm is used to analyze the source domain group, and BWPA (Bayesian Weighted Probability) is used to obtain the probability that the source domain can characterize the target domain, which provides a valuable reference for the optimization method in the following prediction model (the inner layer of the Metas algorithm below needs to be weighted using this probability value).

[0043] Specifically, the improved meta-learning training policy is as follows:

[0044] Meta-learning optimizes the model's initialization parameters by performing gradient descent updates for each task and minimizing the test loss across all tasks using a meta-optimizer, enabling rapid adaptation to new tasks and improving the generalization ability of the predictive model. While Tnet effectively solves the key issue of source domain selection in meta-learning and meets the strict requirements for high-quality prior tasks, it still has limitations in its parameter update mechanism. Specifically, meta-layer updates optimize the initialization parameter state across tasks using inner-layer gradient descent, but this design can lead to double contradictions in time series prediction tasks. On the one hand, the inner-layer updates fail to capture the dynamic correlations of time series features, making it difficult for model parameters to adequately adapt to complex modes in the time dimension. On the other hand, shared initialization parameters across tasks lack adaptive adjustment to source domain differences during cross-domain transfer, ultimately resulting in a significant decline in the model's prediction accuracy. Drawing on the insights of meta-learning, we propose an improved meta-learning parameter optimization method. As shown in Figure 3, three source domains are selected as representatives, and θ (0)represents the initial parameters in the prediction model, and θ (m+1) represents the final parameters obtained after m+1 updates of the prediction model, where x1, x2, and x3 represent the probability of a certain complex being the source domain, L1 represents the loss corresponding to complex 1, ∇L1 performs the gradient update, and α∇L1 represents the inner update step size. Similarly, α∇L1 and α∇L1 are the parameters corresponding to complex 2 and complex 3, respectively. The Metas algorithm uses two-layer updates (including inner and outer layers) to affect all of the model's hyperparameters. The inner layer contains the loss values ​​of each task and the source domain weight information calculated by Tnet, which are dynamically incorporated into the parameter update of the outer layer to achieve differentiated gradient adjustment. The loss values ​​and weight values ​​of each task in the inner layer are fully applied to the update function of the outer layer, and a dynamic gradient weighting mechanism is used to achieve "inner layer loss-outer layer update." This fully takes into account the parameter updates of different source domains and improves prediction accuracy. This is shown in Equation (6). where θ is the model vector parameter,

number

number

number

[0045]

number

[0046] Based on the above analysis, this application proposes a multi-step energy load forecasting model (Metas-Tnet-CNN-MHA-LSTM) based on transfer learning. As shown in Figure 4, the model is an m2m multi-step load forecasting model. In Figure 4, i (1) ~i (n) represents the input, and o (1) ~o (n) represents the predicted output, and y represents the actual value used to calculate the loss along with the predicted value. t where t represents the time step and h t represents the output vector of the long-term and short-term time network at a certain time, which is transmitted to the multi-head attention mechanism (MHA). The principle of MHA is to analyze the local features of the time series and find the actual content weighted by focusing on the features. t into two parts (h' t The reason for dividing it into h t MHA analysis was performed on the upper h t For α t Weighted by the lower h t α t Here, n is the head of the MHA, i.e., the head of the multi-head attention mechanism.

[0047] Figure 4(a) shows the overall model framework. Let x be the number of housing complexes in the source domain, m be the number of features, and n be the number of output features. The forecasting model has an m2m structure and consists of an encoder module and a decoder module, as shown in Figure 4(b). The input is different weather features, and the output is the cooling, heating, electricity, and gas forecast values. The encoder uses a CNN layer to perform feature extraction and dimensionality reduction through convolutional operations, and the compressed data is then sent to an LSTM to further capture long-term dependencies in the time series. A multi-head attention mechanism is used to dynamically weight the relationships between different time steps, assigning different hidden state probability weights to the LSTM, thereby focusing attention on the influence of important information related to cooling, heating, electricity, and gas loads. This further optimizes the model's global modeling capability. Finally, the hidden layer relationships obtained by the encoder module are input to a decoder module, which is based on an LSTM, to output the final load forecast values.

[0048] Thus, a multi-energy load forecasting model without historical load data based on transfer learning is established. To verify the effectiveness of the proposed method, three experimental examples are carried out to illustrate it.

[0049] Example 1 is a forecast for a commercial complex.

[0050] As source domain data, electricity, heating, gas, and cooling load data (sampled every hour) from nine housing complexes and six meteorological features (temperature, humidity, wind speed, etc.) were selected, and normalization and outlier processing were performed. As target domain data, only six meteorological data were input, and historical load records were not required. The Tnet algorithm was run to select appropriate source domains, which were 1, 7, and 8. Comparative experiments were then conducted using other source domains. The results are shown in Table 1.

[0051] [Table 1]

[0052] As can be seen from the experimental data in Table 1, the Tnet source domain selection algorithm demonstrates significant advantages in source domain selection. Comparing the Metas-CNN-MHA-LSTM model with different source domains clearly reveals the performance differences between each source domain and different target domains. In particular, for target complex 0, the Tnet algorithm successfully selected source domain groups 1, 7, and 8 as the optimal source domain groups, significantly improving the model's fitness and prediction accuracy. Further analysis revealed that source domains 5 and 9 had the lowest prediction effectiveness, which is entirely consistent with the low probability of Tnet selecting 5 and 9. Next, source domains 1, 4, 7, and 8 had slightly lower prediction effectiveness, which is due to the lower accuracy caused by the probability of including source domain 4 (<80%). The reason for not selecting the two source domains with the highest probabilities is that we want to fully utilize the three source domains with the highest probabilities for prediction. The similarity of their probabilities and the similarity of the final prediction results further demonstrate the reliability of Tnet's output probability values. These experimental analyses show that the advantage of the Tnet algorithm is that it can accurately select the optimal source domain group in different source and target domain combinations, helping the model achieve better performance. Especially in tasks with multiple source and target domains, the Tnet algorithm clearly has high adaptability and effectiveness optimization capabilities.

[0053] Example 2 is a forecast for an administrative complex.

[0054] To further verify the generalization ability of the proposed prediction model across different scenarios, this experiment conducted cross-domain and cross-application scenario tests. The same model architecture, Metas-Tnet-CNN-MHA-LSTM, was adopted, and the model's adaptability across different scenarios was analyzed to verify its cross-scene robustness. An office complex was selected as a new target domain. Its cooling load is significantly affected by weekday periodicity, forming a stark contrast with the commercial complex discussed in Section 3.4. The Tnet algorithm screened out the most probable source domain groups (1, 5, 9) for the office complex. According to the results in Table 2, when the preferred source domain group (5, 9) was adopted, the MAPEs for electricity, heating, and cooling loads were reduced to 10.166%, 10.341%, and 9.342%, respectively, resulting in a significant reduction in R 2 All of the values ​​were above 85%, significantly better than the other combinations. When low-probability source domains (e.g., 2, 4) were included, the cooling load prediction R 2 It should be noted that the value even becomes negative (-45.2%), indicating that heterogeneous feature interference causes model failure. Tnet uses Bayesian weighting to eliminate such interfering source domains (with probability <60%), ensuring that the model focuses on subgroups with high generalization ability. In contrast, the cooling load MAPE of traditional single-source domain transfer (e.g., using only Apartment 5) is 12.557%, an error 34.4% higher than Tnet's preferred combination, thereby highlighting the need for multi-source domain collaboration. Furthermore, source domain 3 is highly matched with the target domain in time-periodic components and local nonlinear trends, so it is given a higher weight (70%-80% probability) by the Tnet algorithm, and its heating load prediction is considered important information (MAPE = 15.585%, R 2 = 67.7%), demonstrating that Tnet can adaptively adjust the priority of source domains based on multidimensional features.

[0055] The Metas-Tnet-CNN-MHA-LSTM model dynamically captures cross-domain spatiotemporal dependencies through a multi-head attention mechanism, mitigating the distribution shift problem. For example, in gas load forecasting, the linear correlation between the target domain and source domains 5 and 9 is lower than in other load forecasts. However, the model further extracts cross-domain nonlinear dependencies (e.g., load fluctuation modes) through attention weights, resulting in a significantly lower MAPE (16.421%) than the overall source domain combination (19.372%). Furthermore, the Metas optimization policy considers the adaptability of different source domain features to the target domain during the gradient update process, allowing the model to learn optimal parameter settings more quickly and further improve prediction accuracy. In similar scenarios, the Tnet algorithm effectively recognizes groups of source domains with heterogeneous distributions but complementary functions through spatiotemporal feature decoupling and optimal probability selection. Metas-Tnet-CNN-MHA-LSTM achieves directional enhancement of cross-domain nonlinear features through attention mechanism, and the collaboration between them results in a high accuracy in similar generalization tasks (MAPE<10.15%, R 2 >83.7%). Furthermore, the experimental verification in this section demonstrates that the proposed model can not only be applied to specific target domains, but can also be disseminated to different application scenarios, providing a theoretical basis for the standardization of integrated energy systems at the housing complex level.

[0056] [Table 2]

[0057] Example 3 is a small sample fine-tuning optimization prediction.

[0058] Based on the data-free transfer results described above, to further explore the effect of fine-tuning a small amount of target domain data on the prediction model, this section introduces a small-sample transfer learning policy and analyzes the impact of different fine-tuning data ratios on model accuracy. In the experiments, we fine-tuned the model using 3.3% (12 h), 10% (36 h), and 16.7% (60 h) of target domain data. We also used a layer-wise freezing strategy (LFS), i.e., we froze the first two layers of the CNN and LSTM, and only thawed the final layer of the LSTM and the multi-head attention layer. This not only preserves the common features learned by the source domain, but also adapts to the specific characteristics of the target domain, improving prediction accuracy. Table 3 shows the prediction results with different amounts of transfer data.

[0059] Table 3 shows that after introducing 3.33% of the target domain data (12 hours), the cooling load MAPE decreased from 9.329% to 7.094%, and R 2 indicates a 6.9% increase to 85.5%. When the refinement data increases to 16.7% (60 hours), the electrical load MAPE increases to 7.916% (R 2 = 89.6%), and the error is reduced by 13.2% compared to the case without samples. This improvement is due to the dynamic adjustment of the attention layer. That is, by updating the multi-head attention weights, the model strengthens feature extraction of important time steps in the target domain (e.g., weekday peaks) and effectively mitigates local distribution differences. Gas Load Prediction R 2 The value rises to 86.1% after fine-tuning, indicating that the model can quickly calibrate the feature mapping relationship with a small amount of data. It should be noted that while a small amount of data (e.g., 3.3%) can significantly improve the prediction accuracy, the improvement tends to converge as the amount of data increases further (e.g., 16.7%).

[0060] This is because the prediction model based on the Tnet-Metas framework already has strong generalization capabilities and can fully utilize source domain knowledge to make accurate predictions even when the amount of fine-tuning data is small. Experimental results also verified the effectiveness of the stratified freezing strategy in capturing rules specific to the target domain by retaining global features from the source domain and adjusting local parameters. This avoids the problem of overfitting caused by excessive adjustment of lower-level feature extraction layers during the fine-tuning process. This method has significant application value in scenarios where incremental data acquisition is possible.

[0061] [Table 3]

[0062] Example 2 In this embodiment, a data acquisition and preprocessing module configured to acquire meteorological characteristics of the target complex and historical data of cooling, heating, electricity, and gas of the source domain group complex, and preprocess the acquired data; a source domain data determination module configured to perform a cross-correlation and generalization ability analysis of the cooling, heating, electricity, and gas historical data of the source domain group housing complexes after preprocessing to determine appropriate source domain data; a model building and training module, configured to build a multi-energy load forecasting model, train the model based on source domain data according to a Metas training policy, and adjust gradient weights based on the validation loss and source domain generalization probability of the inner-loop task to obtain a trained forecasting model; a model output module configured to input the preprocessed meteorological characteristics of the target area into a forecasting model to obtain a forecast result; The present invention provides a multi-energy load forecasting system for an integrated energy system without historical data, including:

[0063] It should be noted that the above modules correspond to the steps described in Example 1, and the examples and application scenarios in which the above modules and corresponding steps are realized are the same, but are not limited to those disclosed in Example 1. It should be noted that the above modules can be executed in a computer system as a set of computer-executable instructions as part of the system.

[0064] In more embodiments, the following are further provided:

[0065] An electronic device includes a memory, a processor, and computer instructions stored in the memory and executed on the processor, which, when executed by the processor, completes the method described in Example 1. For the sake of brevity, detailed descriptions are omitted here.

[0066] It should be understood that in this embodiment, the processor may be a central processing unit CPU, or other common processor, a digital signal processor DSP, an application specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, etc. The common processor may be a microprocessor, or the processor may be any common processor, etc.

[0067] The memory may include read-only memory and random access memory to provide instructions and data to the processor, and a portion of the memory may further include non-volatile random access memory. For example, the memory may further store device type information.

[0068] A computer-readable storage medium used to store computer instructions, which, when executed by a processor, complete the method described in Example 1.

[0069] The method in Example 1 can be directly implemented by a hardware processor, or by a combination of hardware and software modules in a processor. The software modules may be stored in a storage medium commonly used in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the method in combination with the hardware. Detailed descriptions are omitted here to avoid repetition.

[0070] A computer program product includes a computer program, which, when executed by a processor, accomplishes the method described in Example 1.

[0071] The present invention further provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, e.g., instructions included in program modules, that execute on a target device on a real or virtual processor to perform the process / method described above. Typically, program modules include routines, programs, libraries, objects, types, components, data structures, etc. that perform particular tasks or implement particular abstract data types. In various embodiments, the functionality of the program modules may be combined or divided among program modules as desired. Machine-executable instructions utilized in program modules may be executed in local or distributed devices. In a distributed device, program modules may be located in both local and remote storage media.

[0072] Computer program code for implementing the methods of the present invention can be written in one or more programming languages. These computer program codes may be provided to a processor of a common computer, a special purpose computer, or other programmable data processing apparatus, so that when the program code is executed by the computer or other programmable data processing apparatus, the functions / operations specified in the flowcharts and / or block diagrams are performed. The program code can be executed entirely on the computer, partially on the computer, as a separate software package, partially on the computer and partially on a remote computer, or entirely on a remote computer or server.

[0073] In the context of the present invention, computer program code or associated data may be carried on any suitable carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, etc. Examples of signals include electrical, optical, wireless, acoustic or other forms of propagated signals, such as carrier waves, infrared signals, etc.

[0074] It is understood by those skilled in the art that the units and algorithm steps of each example described in this embodiment can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in the form of hardware or software is determined by the specific application and design constraints of the technical solution. Professionals can implement the described functions using different methods for each specific application, but it should not be understood that such implementation goes beyond the scope of this application.

[0075] Although the specific embodiments of the present invention have been described above with reference to the drawings, they are not intended to limit the scope of protection of the present invention. As those skilled in the art will appreciate, various modifications or variations made by those skilled in the art based on the technical solutions of the present invention without any creative efforts will still fall within the scope of protection of the present invention.

Claims

1. Obtaining meteorological characteristics of the target complex and historical data of cooling, heating, electricity, and gas of the source domain group complex, and preprocessing the obtained data; A step of performing cross-correlation and generalization ability analysis of the cooling, heating, electricity, and gas history data of the source domain group housing complex after preprocessing to determine appropriate source domain data; The cross-correlation and generalization ability analysis of the complex, specifically: Using the Time2vec algorithm to convert multiple load time series data of each source domain into two parts, periodic and linear, and processing them respectively using the MIC algorithm and the MD algorithm; constructing a K-means cluster analysis to perform generalized potential hierarchy for the source domain, removing low potential clusters to reduce noise interference; and performing weighted fusion for MD and MIC; using grid search to find the optimal parameter range; obtaining the cross-correlation score of the source domain group and the reasonable source domain range; and obtaining the probability that the source domain can characterize the target domain through Bayesian weighted probability; Construct a multi-energy load forecasting model, train the model based on source domain data according to the Metas training policy, and adjust the gradient weights based on the validation loss of the inner-loop task and the source domain generalization probability to obtain a trained forecasting model; inputting the pre-processed weather characteristics of the target area into a forecasting model to obtain a forecast result; A computer-implemented method for multi-energy load forecasting in the absence of historical data for an integrated energy system, comprising:

2. Specifically, the Metas training policy: The method for multi-energy load forecasting without historical data of an integrated energy system according to claim 1, characterized in that the parameter updates of different source domains are taken into consideration, and the loss values ​​and weight values ​​of each training task in the inner layer are fully applied to the update function of the outer layer.

3. The method of claim 1 , wherein the multi-energy load forecasting model comprises an encoder module and a decoder module.

4. The method for multi-energy load forecasting without historical data of an integrated energy system according to claim 3, wherein the encoder uses a CNN layer to perform feature extraction and dimensionality reduction through a convolutional operation, and transmits the compressed data to an LSTM to further capture long-term dependencies in the time series.

5. The method for multi-energy load forecasting without historical data of an integrated energy system according to claim 4, characterized in that by utilizing a multi-head attention mechanism to dynamically weight the relationships between different time steps, different hidden state probability weights are assigned to the LSTM, and attention is focused on the influence of important information related to cooling, heating, electricity, and gas loads.

6. a data acquisition and pre-processing module configured to acquire meteorological characteristics of the target complex and cooling, heating, electricity, and gas history data of the source domain group complex, and pre-process the acquired data; A source domain data determination module is configured to perform a cross-correlation and generalization ability analysis of the cooling, heating, electricity, and gas historical data of the source domain group housing complex after preprocessing to determine appropriate source domain data, The cross-correlation and generalization ability analysis of the complex, specifically: a source domain data determination module that uses the Time2vec algorithm to transform multiple load time series data of each source domain into two parts, periodic and linear, and processes them respectively using the MIC algorithm and the MD algorithm; constructs a K-means cluster analysis to perform generalized potential hierarchy for the source domain, removes low-potential clusters to reduce noise interference, and performs weighted fusion for MD and MIC; uses grid search to find the optimal parameter range, obtains the cross-correlation score of the source domain group and the reasonable source domain range; and obtains the probability that the source domain can characterize the target domain through Bayesian weighted probability; a model building and training module configured to build a multi-energy load forecasting model, train the model based on source domain data according to a Metas training policy, and adjust gradient weights based on a validation loss of an inner-loop task and a source domain generalization probability to obtain a trained forecasting model; a model output module configured to input the preprocessed meteorological characteristics of the target area into a forecasting model to obtain a forecast result; A multi-energy load forecasting system for an integrated energy system without historical data, comprising:

7. 6. An electronic device comprising: a memory; a processor; and computer instructions stored in the memory and executable on the processor, the computer instructions, when executed by the processor, completing the method of any one of claims 1 to 5.

8. A computer-readable storage medium used to store computer instructions that, when executed by a processor, complete the method of any one of claims 1 to 5.

Images