Method and system for predicting multi-energy loads in the absence of historical data for integrated energy systems
Patent Information
- Application Number
- JP2025132272
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-04-03
- Filing Date
- 2025-08-07
- Publication Date
- 2026-10-16
- Estimated Expiration
- 2045-08-07
Smart Images
Figure 2026169114000001_ABST
Abstract
Description
[Technical Field]
[0001] (Cross-reference of related applications) This invention claims priority to the Chinese patent application filed with the China National Intellectual Property Administration on April 3, 2025, application number 202510412447.8, with the title "Multi-Energy Load Prediction Method and System for Integrated Energy Systems in the Case of No Historical Data," the entire contents of which are incorporated by reference into this invention and constitute part of this invention, and are used for all purposes.
[0002] The present invention relates to the technical field of load prediction, and more particularly to a multi-energy load prediction method and system for integrated energy systems in the absence of historical data. [Background technology]
[0003] This statement is solely for the purpose of providing background information related to the present invention and does not necessarily constitute prior art.
[0004] With the rapid development of Integrated Energy Systems (IES), accurate prediction of multi-energy loads has become a critical requirement for optimizing system operation. However, in practical applications, obtaining comprehensive multi-energy load historical data generally requires significant time and financial resources. Furthermore, due to data privacy concerns, some information cannot be made public, limiting the size of available datasets. Conventional deep learning methods face serious data dependency bottlenecks, posing a significant challenge to IES load prediction. In response, many scholars both in China and abroad are employing transfer learning methods to address this challenge.
[0005] Performing correlation analysis on loading sequences to assist in selecting appropriate source domains is one of the most crucial steps before transfer learning. Some scholars consider the complex linear and nonlinear features of the loading data, thereby improving the correlation analysis and further reducing the amount of target domain data required. However, in extreme situations where no housing complex has been built, the target system may not even have any historical loading data. Traditional source domain selection relies on target domain correlation analysis, but feature comparison is completely useless under data-free conditions. Some scholars propose transfer learning strategies based on Generative Adversarial Networks (GANs), fine-tuning policies, and meta-learning frameworks, but these still require small amounts of target domain data and have clear shortcomings in terms of extracting complex time-series features and coordinating optimization of multi-source domains.
[0006] As described above, scholars both in China and abroad have proposed various solutions for high-accuracy multi-energy load prediction under conditions where load data is extremely scarce. However, under conditions where historical load data for newly built residential areas is unavailable, conventional transfer learning techniques face a double bottleneck.
[0007] (1) Selecting a source domain is difficult. In transfer learning, the similarity analysis between the source domain and the target domain is highly dependent on the target domain data, so conventional correlation analysis methods are ineffective under conditions where there is no data.
[0008] (2) Insufficient cross-domain generalization ability. Cross-domain parameter transfer does not adequately consider the dynamic variability of multi-source domains, resulting in reduced parameter transfer efficiency and larger prediction errors. [Overview of the Initiative]
[0009] To solve the above problems, the present invention proposes a multi-energy load prediction method and system for integrated energy systems in the absence of historical data, constructs a transfer learning-based multi-energy load prediction model based on the Tnet algorithm and an improved meta-learning policy, and achieves accurate long-term prediction of multi-energy loads in a target integrated energy system under the condition of no historical load data.
[0010] To achieve the above objective, the present invention employs the following technical solutions.
[0011] In the first aspect, the present invention is The steps include acquiring weather characteristics of the target housing complex and historical data on air conditioning, heating, electricity, and gas for the source domain group housing complex, and preprocessing the acquired data. The following steps are taken to determine appropriate source domain data by performing cross-correlation and generalization capacity analysis on the pre-processed historical data of air conditioning, heating, electricity, and gas for the source domain group housing complexes. The steps include constructing a multi-energy load prediction model, training the model based on source domain data using the Metas training policy, adjusting gradient weights based on the validation loss of the internal loop task and the source domain generalization probability to obtain a trained prediction model, and The steps include inputting the pre-processed weather characteristics of the target housing complex into a prediction model and obtaining prediction results, This provides a method for predicting multi-energy loads in the absence of historical data for integrated energy systems.
[0012] As an optional embodiment, the Metas training policy specifically: The goal is to fully apply the loss and weight values of each training task in the inner layers to the update function of the outer layers, taking into account parameter updates from different source domains.
[0013] In a selective embodiment, the multi-energy load prediction model comprises an encoder module and a decoder module.
[0014] As an optional embodiment, the encoder uses CNN layers to perform feature extraction and dimension reduction through convolution operations, and transmits the compressed data to the LSTM to further capture long-term dependencies in the time series.
[0015] As an optional embodiment, by using a multi-head attention mechanism to dynamically weight the relationships between different time steps, different hidden state probability weights are assigned to the LSTM, so as to focus attention on the influence of important information related to cooling, heating, electricity and gas loads.
[0016] As an optional embodiment, the cross-correlation and generalization capability analysis of the residential complex specifically includes: Using the Time2vec algorithm to convert multiple pieces of load time series data of each source domain into two parts: periodicity and linearity, processing them using the MIC algorithm and the MD algorithm respectively, constructing K-means cluster analysis to perform generalization potential stratification on the source domains, removing low-potential clusters to reduce noise interference, performing weighted fusion on MD and MIC, using grid search to find the optimal parameter range, and obtaining the cross-correlation score of the source domain group and a reasonable source domain range.
[0017] In a second aspect, the present invention relates to: a data acquisition and preprocessing module, which is configured to acquire meteorological features of the target residential complex and cooling, heating, electricity and gas historical data of the source domain group residential complexes, and preprocess the acquired data; a source domain data determination module, which is configured to perform cross-correlation and generalization capability analysis of the residential complexes on the preprocessed cooling, heating, electricity and gas historical data of the source domain group residential complexes, so as to determine appropriate source domain data; A model building and training module is configured to construct a multi-energy load prediction model, train the model based on source domain data using the Metas training policy, adjust gradient weights based on the validation loss of the internal loop task and the source domain generalization probability, and obtain a trained prediction model. A model output module is configured to input the pre-processed weather characteristics of the target housing complex into a prediction model and obtain prediction results. This provides a multi-energy load prediction system for integrated energy systems, including the case where historical data is unavailable.
[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 executed on the processor, wherein the method according to the first aspect is completed when the computer instructions are executed by the processor.
[0019] In a fourth aspect, the present invention provides a computer-readable storage medium used for storing computer instructions, wherein the method described in the first aspect is completed when the computer instructions are executed by a processor.
[0020] In a fifth aspect, the present invention provides a computer program product that includes a computer program, and when the computer program is executed by a processor, it realizes the completion of the method described in the first aspect.
[0021] Compared to conventional technology, the beneficial effects of the present invention are as follows:
[0022] The present invention provides a multi-energy load prediction method and system for integrated energy systems in the absence of historical data. Based on the Tnet algorithm and an improved meta-learning policy, it constructs a multi-head attention optimized coding-decoding model using transfer learning, achieves breakthrough predictions through three-stage cascade optimization, realizes long-term predictions for cooling, heating, electricity, and gas loads of a target domain, and significantly improves load prediction accuracy in the absence of data.
[0023] The present invention provides a multi-energy load prediction method and system for integrated energy systems in the absence of historical data. This method proposes a source domain selection algorithm (Tnet) for the absence of historical load data for target domains. It transforms conventional source domain similarity analysis into a model generalization ability probabilistic evaluation task. Without requiring any historical target domain data, Tnet screens for the most appropriate set of source domains, enabling load prediction and overcoming the initialization challenge of transfer learning under the constraint of no data.
[0024] The present invention provides a multi-energy load prediction method and system for integrated energy systems without historical data, which incorporates an improved meta-learning training algorithm (Metas) that introduces loss values and weight information for each inner loop task when updating the outer loop of meta-learning, fully utilizes multi-source domain knowledge provided by Tnet, overcomes parameter transition barriers in time series prediction, and improves the generalization performance of the prediction model.
[0025] Additional advantages of the present invention are, in part, set forth in the following description, in part, become apparent from the following description, or are understood through the implementation of the present invention.
[0026] The drawings in the specification, which constitute part of the present invention, are for further understanding of the present invention, and the exemplary embodiments and descriptions thereof are for interpretation purposes only and are not intended to improperly limit the present invention. [Brief explanation of the drawing]
[0027] [Figure 1] This is a flowchart of the multi-energy load prediction method for an integrated energy system provided in Embodiment 1 of the present invention, in the case of no historical data. [Figure 2] This is a flowchart of the TNet algorithm provided in Example 1 of the present invention. [Figure 3] This is a schematic diagram of the Metas algorithm provided in Embodiment 1 of the present invention. [Figure 4] Figure 4(a) is a framework diagram of the multi-energy load prediction model provided in Embodiment 1 of the present invention, where Figure 4(a) is an overall framework diagram of the model and Figure 4(b) is a framework diagram of the model's coder and decoder. [Modes for carrying out the invention]
[0028] The present invention will be further described below with reference to the drawings and embodiments.
[0029] It should be noted that the following detailed descriptions are illustrative and intended to further illustrate the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art.
[0030] It should be noted that the terms used herein are for the sole purpose of describing specific embodiments and are not intended to limit the exemplary embodiments of the present invention. For example, unless otherwise specified in the context, the singular form used herein is intended to include plural forms, and furthermore, the terms “includes,” “has,” and any variations thereof are intended to cover non-exclusive inclusion, and it should be further understood that, for example, a process, method, system, product, or apparatus comprising a series of steps or units is not limited to the steps or units expressly listed, and may include other steps or units not expressly listed or specific to these processes, methods, products, or apparatus.
[0031] The embodiments and features of the present invention can be combined with each other without contradiction.
[0032] Example 1 As shown in Figure 1, this embodiment is The steps include acquiring weather characteristics of the target housing complex and historical data on air conditioning, heating, electricity, and gas for the source domain group housing complex, and preprocessing the acquired data. The following steps are taken to determine appropriate source domain data by performing cross-correlation and generalization capacity analysis on the pre-processed historical data of air conditioning, heating, electricity, and gas for the source domain group housing complexes. The steps include constructing a multi-energy load prediction model, training the model based on source domain data using the Metas training policy, adjusting gradient weights based on the validation loss of the internal loop task and the source domain generalization probability to obtain a trained prediction model, and The steps include inputting the pre-processed weather characteristics of the target housing complex into a prediction model and obtaining prediction results, This provides a method for predicting multi-energy loads in the absence of historical data for integrated energy systems.
[0033] To address the extreme problem of the lack of historical load data for target domains, this application proposes a brand-new multi-energy load prediction method. The method includes an encoded-decoded prediction model that combines a novel source domain selection algorithm, an improved meta-learning training policy, and a multi-head attention mechanism, achieving prediction breakthroughs through three-stage cascade optimization, thereby enabling accurate long-term predictions for cooling, heating, electricity, and gas loads in a target integrated energy system. It is applicable to the planning and scheduling of newly constructed housing complexes or integrated energy systems lacking historical data.
[0034] As shown in Figure 1, first, preprocessing is performed on the raw data of the source domain group and target housing complex, including anomaly detection, data normalization, and feature contribution assessment. Next, the Tnet algorithm is used to perform cross-correlation and generalization ability analysis of housing complexes in the source domain group, transforming the similarity analysis problem into a classification problem and finding housing complexes suitable as source domains. Subsequently, autocorrelation analysis is constructed by combining this with the data preprocessing of the target housing complex to determine the type and period of the input data. Then, data from multiple housing complexes are input into the prediction model, and features are extracted sequentially using a Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Multi-Head Attention (MHA) encoding / decoding module to decompose and extract combined information and obtain joint prediction results. Metas is used to apply the loss values and weight values of each task in the inner layer to the function in the outer layer, and the model parameters are further updated. Finally, the weather features of the target housing complex are added to the updated model to realize multi-energy prediction of cooling, heating, electricity, and gas load data in the case of no historical data.
[0035] Here, using the Tnet algorithm to perform cross-correlation and generalization ability analysis of housing complexes for a source domain group, and to find suitable housing complexes as source domains, is done specifically as follows:
[0036] The Tnet algorithm reconstructs the source domain selection paradigm through a five-step probabilistic inference framework, finding appropriate source domain data for time vector embedding (Time to Vector, Time2vec), Mahalanobis distance (MD), Maximal Information Coefficient (MIC), K-means clustering (K-means), and Bayesian Weighted Probability Averaging Method (BWPA) through five main steps. In Figure 1,
number
number
number
[0037] Figure 2 shows the flowchart. Since the historical data for air conditioning, heating, electricity, and gas loads of the target housing complex is unknown, it is crucial to find a group of similar housing complexes as source domains from the source domain group. Given the periodicity and linear features of the load data used, the single consideration of linear and nonlinear elements, and the direct use of linear and nonlinear analysis, will inevitably be affected by opposing factors. Therefore, the Time2vec algorithm is used to transform multiple load time series data from each source domain into two parts: periodicity and linearity, and to map the time features into a single high-dimensional space. Periodicity mainly considers seasonal changes in load prediction, while linear features capture the baseline trend of the load data. This is shown in equation (1). In the equation, t2v(τ) represents the initial time series, τ represents the time step width, i represents the embedding dimension, 0 dimensions represent linear features, and 1 to k dimensions represent nonlinear features.
[0038]
number
[0039] To perform joint linear and nonlinear analysis in a high-dimensional space, the MD algorithm is used to introduce a covariance matrix, and the weights of different features are adjusted, so that features of different scales and different correlations are treated fairly during computation. The MIC algorithm is used to divide different grids for different nonlinear variables, and the maximum amount of information is found through different discretization policies, as shown in Equations (2) and (3). Furthermore, K-means cluster analysis is constructed using MD and MIC to perform generalization potential stratification on the source domain, and low-potential clusters are removed to reduce noise interference. One cluster contains multiple building complexes, and if they are used in a unified manner, the respective reliable generalization ability of each building complex relative to the target building complex cannot be considered, resulting in a very large prediction error. Therefore, weighted fusion is performed on MD and MIC according to Equation (4), and a grid search is used to find the optimal parameter range, so as to obtain the cross-correlation score of the source domain groups in this cluster and a reasonable source domain range.
[0040] [Equation]
[0041] Finally, in order to make full use of the contribution of different source domains to the target domain in the prediction model, weighted mutual information entropy is introduced to construct a Bayesian weighted model, reasonable source domains are taken as tags, a parameter model is constructed with α in Equation (4) and substituted into Equation (5), and weights are assigned by considering the "uncertainty" of each data point. In this way, the conventional probabilistic prediction problem is transformed into a classification problem, and the most appropriate building complex group is found. Here, X and Y are the cooling, heating, electricity, and gas data of two different building complexes, both of which are 4-dimensional. Σ -1 represents the inverse matrix of Σ, which denotes the correlation of each feature between data points and is a 4×n matrix. G refers to dividing the data into an a×b grid, I(X,Y|G) is the mutual information estimate on grid G, and the maximum value is selected from all grids satisfying a×b<B(n) (usually, B(n)=n 0.6(where n is the sample size), P(M i ) is Model M i Here, y is the prior probability, y is the predicted result, and X is the input feature.
[0042] Therefore, in this application, this algorithm can be viewed as a problem of maximizing the generalization ability after standardizing data from four different scales: cooling, heating, electricity, and gas. At the same time, the Tnet algorithm is used to analyze the source domain group, and the probability that the source domain can characterize the target domain is obtained by BWPA (Bayesian Weighted Probability), which also provides a valuable reference for the optimization method in the following prediction model (the inner layers of the Metas algorithm below need to be weighted using these probability values).
[0043] 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 for all tasks using a meta-optimizer, thereby enabling rapid adaptation to new tasks and improving the generalization ability of the predictive model. Tnet effectively solves a key problem in source domain selection in meta-learning and meets the stringent requirements of meta-learning for high-quality pre-tasks, but it still has limitations in terms of parameter update mechanisms. Specifically, meta-layer updates optimize the initialization parameter state between tasks by inner layer gradient descent, and such a design can lead to double inconsistencies in time series prediction tasks. On the one hand, the lack of dynamic correlation capture of time series features by inner layer updates makes it difficult for model parameters to adequately fit complex modes in the time dimension, and on the other hand, initialization parameters shared between tasks lack adaptive adjustment for source domain differences during cross-domain transitions, ultimately leading to a significant decay in the model's prediction accuracy. Inspired by the meta-learning philosophy, we propose an improved meta-learning parameter optimization method. As shown in Figure 3, three source domains are selected as representatives, and θ (0)denotes the initial parameter in the prediction model, and θ (m+1) denotes the final parameter obtained after the prediction model undergoes m+1 updates, where x1, x2~x3 represent the probabilities of a certain residential estate being set as the source domain, L1 represents the loss corresponding to residential estate 1, ∇L1 represents gradient update, α∇L1 represents the inner update step size, and similarly, α∇L1 and α∇L1 are parameters corresponding to residential estate 2 and residential estate 3, respectively. The Metas algorithm applies all two-layer updates to the hyperparameters of the model through two-layer updates (including an inner layer and an outer layer). The inner layer is the loss value of each task and the source domain weight information calculated by Tnet, which is dynamically incorporated into the parameter update of the outer layer to realize differentiated gradient adjustment. The loss value and weight value of each task in the inner layer are fully applied to the update function of the outer layer, so as to realize "inner layer loss - outer layer update" through the dynamic gradient weighting mechanism. In this way, the parameter updates of different source domains can be fully considered, and the prediction accuracy can be improved, as shown in Equation (6). In the equation, θ is the model vector parameter, and [Math] is the task T i training loss parameter, and [Math] is the task T i validation loss parameter, ω i is the task weight calculated by the Tnet algorithm, α is the inner gradient update step size, β is the outer layer learning rate, B is the batch size, and f θ (x) is the model prediction output of the input x, and [Math] is the Hessian matrix of the task T i training loss.
[0045] [Math]
[0046] Based on the above analysis, this application proposes a transfer learning-based multi-energy load prediction model (Metas-Tnet-CNN-MHA-LSTM). As shown in Figure 4, the model is an m2m multi-step load prediction model. In Figure 4, i (1) ~i (n) represents the input, o (1) ~o (n) represents the predicted output, and y represents the actual value used to calculate the loss along with the predicted value. h is the final output from the LSTM (Long Short-Term Memory Network). t In this, t represents a time step, and h t This represents the output vector of the long- and short-term time network at a given time, which is transmitted to the Multi-Head Attention Mechanism (MHA). The principle of the MHA is to analyze the local features of the time series, focus attention on those features, and find weighted actual content. t two parts (h' t The reason for splitting it into h t MHA analysis was performed on the upper h t For α t Weighting is applied by the following, and the lower h t to α t This is because it is used in the calculation of [the value]. Here, n is the head of the MHA, that is, the head of the multi-head attention mechanism.
[0047] Here, Figure 4(a) shows the overall framework of the model, where x is the number of housing complexes in the source domain, m is the number of features, and n is the number of output features. The prediction model has an m2m structure and consists of an encoder module and a decoder module, as shown in Figure 4(b). The inputs are different weather features, and the outputs are predicted values for cooling, heating, electricity, and gas. The encoder uses a CNN layer to perform feature extraction and dimensionality reduction through convolutional operations, and the compressed data is transmitted to the LSTM to further capture long-term dependencies in the time series. By dynamically weighting the relationships at different time steps using a multi-head attention mechanism, different hidden state probability weights are assigned to the LSTM, drawing attention to the influence of important information related to cooling, heating, electricity, and gas loads. This further optimizes the model's global modeling capabilities. Finally, the hidden layer relationships obtained in the encoder module are input to the decoder module, which has an LSTM as its core, to output the final load prediction values.
[0048] This establishes a multi-energy load prediction model based on transfer learning without historical load data. To verify the effectiveness of the proposed method, three experimental examples are presented for explanation.
[0049] Example 1 is a prediction for a commercial park.
[0050] For the source domain data, we selected electricity, heating, gas, and cooling load data (sampled every hour) from nine housing complexes and six meteorological features (temperature, humidity, wind speed, etc.), and performed normalization and outlier handling. For the target domain data, we input only the six meteorological data points and did not require historical load records. We ran the Tnet algorithm to select appropriate source domains, which were 1, 7, and 8. Subsequently, we conducted comparative experiments 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 shows a significant advantage in source domain selection. By comparing the Metas-CNN-MHA-LSTM model with different source domains, the performance differences exhibited by each source domain for different target domains are clearly visible. In particular, for target housing complex 0, the Tnet algorithm successfully selected source domain groups 1, 7, and 8 as the optimal source domain group, and this selection significantly improves the model's goodness of fit and predictive accuracy. Further analysis revealed that the predictive effect was lowest for source domains 5 and 9, which is perfectly consistent with the result that Tnet has a low probability of selecting 5 and 9. Next, the predictive effect was somewhat low for source domains 1, 4, 7, and 8, which is a decrease in accuracy due to the probability <80% when source domain 4 is included. The reason for not selecting the two source domains with the highest probabilities is that we want to make full use of the three source domains with the highest probabilities for prediction, and the fact that their probabilities are similar and the final prediction results are also sufficiently similar further demonstrates the reliability of Tnet's output probability values. These experimental analyses demonstrate that the superiority of the Tnet algorithm lies in its ability to accurately select the optimal source domain group in combinations of different source and target domains, thereby helping the model achieve better performance. In particular, in multi-source domain and multi-target domain tasks, the Tnet algorithm clearly possesses high adaptability and effectiveness optimization capabilities.
[0053] Example 2 is a prediction for an office complex.
[0054] To further validate the generalization capabilities of the proposed prediction model across different scenarios, this experiment conducted cross-domain and cross-application-scenario testing. The Metas-Tnet-CNN-MHA-LSTM architecture of the same model was employed, and the model's adaptability across different scenarios was analyzed with a focus on its cross-scenario robustness. An office complex was selected as a new target domain. Its cooling load is significantly affected by weekday periodicity, creating a stark contrast with the commercial complex described in Section 3.4. The Tnet algorithm screened high-probability source domain groups (1, 5, 9) for the office complex. According to the results in Table 2, when the preferred source domain groups (5, 9) were selected, the MAPE for electricity, heating, and cooling loads decreased to 10.166%, 10.341%, and 9.342%, respectively. 2 All values were over 85%, significantly better than other combinations. When low-probability source domains (e.g., 2, 4) are included, the cooling load prediction R 2 It should be noted that the values can even be negative (-45.2%), indicating that interference from heterogeneous features can cause model failure. Tnet eliminates such interfering source domains (probability <60%) through Bayesian weighting, ensuring that the model focuses on subgroups with high generalization ability. In contrast, the cooling load MAPE for a conventional single-source domain transition (e.g., using only apartment 5) is 12.557%, and the error is 34.4% higher than that of a favored combination in Tnet, thereby highlighting the need for multi-source domain coordination. Furthermore, source domain 3 is highly matched with the target domain in terms of time-periodic components and local nonlinear tendencies, and is therefore given a higher weight by the Tnet algorithm (probability 70%~80%), and its heating load prediction provides important information (MAPE=15.585%, R 2 This contributes 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 problem of distribution shifts. For example, in gas load prediction, the linear correlation between the target domain and source domains 5 and 9 is lower than in other load predictions, but 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 all source domain combinations (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 the optimal parameter settings more quickly and further improving prediction accuracy. Under similar scenes, the Tnet algorithm effectively recognizes source domain groups with heterogeneous distributions but complementary functions through spatiotemporal feature decoupling and favorable probabilistic selection. Metas-Tnet-CNN-MHA-LSTM achieves directional enhancement of cross-domain nonlinear features through an attention mechanism, and through the cooperation of both, the model achieves high accuracy (MAPE < 10.15%, R) on similar generalization tasks. 2 It maintains >83.7%. Furthermore, the experimental verification in this section shows that the proposed model can be applied not only to specific target domains but also 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 fine-tuning optimization prediction for a small sample.
[0058] Based on the transfer data without any data described above, this section introduces a small-sample transfer learning policy to further explore the effect of fine-tuning the predictive model with small amounts of target domain data, and analyzes the impact of different proportions of fine-tuning data on the model's accuracy. In the experiment, the model was fine-tuned using 3.3% (12h), 10% (36h), and 16.7% (60h) of target domain data, and a layer-wise freezing strategy (LFS) was used, i.e., the first two layers of the CNN and LSTM were frozen, and only the final layer of the LSTM and the multi-head attention layer were unfrozen. This not only preserves the common features learned by the source domain but also adapts to specific modes 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 This indicates an increase of 6.9% to 85.5%. When the fine-tuning data increases to 16.7% (60 hours), the electrical load MAPE is 7.916% (R 2 The model is optimized to 89.6%, and the error is reduced by 13.2% compared to the case without samples. This increase is due to the dynamic adjustment of the attention layer. Specifically, the model enhances feature extraction of critical time steps in the target domain (e.g., weekday peaks) by updating the multi-head attention weights, effectively mitigating local distribution differences. Gas load prediction R 2 The value increased to 86.1% after fine-tuning, indicating that the model can quickly calibrate feature mapping relationships with a small amount of data. It should be noted that while prediction accuracy can be significantly improved with small amounts of data (e.g., 3.3%), the improvement tends to converge as the amount of data increases further (e.g., 16.7%).
[0060] This indicates that the predictive model based on the Tnet-Metas framework already possesses strong generalization capabilities and can make accurate predictions by making full use of source domain knowledge, even with a small amount of fine-tuning data. Furthermore, experimental results have verified the effectiveness of the stratified freezing strategy in capturing rules specific to the target domain by preserving global features of the source domain and adjusting local parameters. This avoids the problem of overfitting due to over-adjustment of the lower-level feature extraction layer 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 This embodiment is, A data acquisition and preprocessing module is configured to acquire weather characteristics of the target housing complex and historical data on air conditioning, heating, electricity, and gas for the source domain group housing complex, and to preprocess the acquired data. A source domain data determination module is configured to perform cross-correlation and generalization capability analysis on the pre-processed historical data of air conditioning, heating, electricity, and gas for the source domain group housing complexes, and to determine appropriate source domain data. A model building and training module is configured to construct a multi-energy load prediction model, train the model based on source domain data using the Metas training policy, adjust gradient weights based on the validation loss of the internal loop task and the source domain generalization probability, and obtain a trained prediction model. A model output module is configured to input the pre-processed weather characteristics of the target housing complex into a prediction model and obtain prediction results. This provides a multi-energy load prediction system for integrated energy systems, including the case where historical data is unavailable.
[0063] What should be explained here is that the above module corresponds to the steps described in Example 1, and the examples and application scenes realized by the above module and the corresponding steps are the same, but are not limited to what is disclosed in Example 1. It should also be explained that the above module can be executed in a computer system as part of a system, such as a set of computer executable instructions.
[0064] Further examples provide the following:
[0065] This electronic device includes memory, a processor, and computer instructions stored in the memory and executed on the processor, wherein the method described in Embodiment 1 is completed when the computer instructions are executed by the processor. For simplicity, a detailed explanation is omitted here.
[0066] In this embodiment, the processor may be a central processing unit CPU, or it may be another 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 this processor may be any general-purpose processor, etc.
[0067] The memory may include read-only memory and random-access memory, providing 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 also store device type information.
[0068] This is a computer-readable storage medium used to store computer instructions, and when the computer instructions are executed by a processor, the method described in Embodiment 1 is completed.
[0069] The method in Example 1 can be directly embodied to be executed and completed by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules may be storage media commonly used in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, or registers. The storage medium is located in memory, and the processor reads the information in memory and combines it with its hardware to complete the steps of the method. A detailed explanation is omitted here to avoid repetition.
[0070] This is a computer program product that includes a computer program, and when the computer program is executed by a processor, it realizes the completion of the method described in Example 1.
[0071] The present invention further provides at least one computer program product tangibly stored in a non-temporary computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions contained in a program module, which are executed on a device on a real or virtual processor of a target and perform the processes / methods described above. Typically, a program module includes routines, programs, libraries, targets, types, components, data structures, etc., that perform a specific task or realize a specific abstract data type. In various embodiments, the functions of program modules can be combined or divided amongst program modules as needed. The machine-executable instructions used in a program module can be executed in a local or distributed device. In a distributed device, program modules can reside in local and remote storage media.
[0072] Computer program code for implementing the method 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 dedicated computer, or another programmable data processing device, so that when the program code is executed by the computer or other programmable data processing device, the functions / operations specified in the flowchart and / or block diagram are performed. The program code can be executed entirely on a computer, partially on a computer, as a standalone software package, partially on a 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 related data may be carried on any suitable carrier so that a device, apparatus, or processor can perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, etc. Examples of signals include electric, optical, wireless telegraph, sound, or other forms of propagating signals, such as carrier waves, infrared signals, etc.
[0074] It will be obvious to those skilled in the art that each example unit and algorithm step described in this embodiment can be implemented using electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software form depends on the specific application and design constraints of the technical solution. While experts in the art may implement the described functions in different ways for each specific application, such implementations should not be understood as exceeding the scope of this application.
[0075] Although specific embodiments of the present invention have been described above with reference to the drawings, this does not limit the scope of protection of the present invention. As those skilled in the art will understand, various modifications or variations made by those skilled in the art based on the technical solutions of the present invention without requiring any creative effort are still included within the scope of protection of the present invention.
Claims
1. The steps include acquiring weather characteristics of the target housing complex and historical data on air conditioning, heating, electricity, and gas for the source domain group housing complex, and preprocessing the acquired data. The step involves performing cross-correlation and generalization capability analysis on the pre-processed historical data of air conditioning, heating, electricity, and gas for the source domain group housing complexes, in order to determine appropriate source domain data. Specifically, the analysis of the intercorrelations and generalization capabilities of housing complexes is as follows: The process involves using the Time2vec algorithm to transform multiple load time series data from each source domain into two parts: periodic and linear. These are then processed using the MIC and MD algorithms, respectively. K-means cluster analysis is constructed to perform generalized potential stratification for the source domains, removing low-potential clusters to reduce noise interference, performing weighted fusion for MD and MIC, finding the optimal parameter range using grid search, obtaining cross-correlation scores and reasonable source domain ranges for the source domain groups, and obtaining the probability that a source domain can characterize a target domain using Bayesian weighted probabilities. The steps include constructing a multi-energy load prediction model, training the model based on source domain data using the Metas training policy, adjusting gradient weights based on the validation loss of the internal loop task and the source domain generalization probability to obtain a trained prediction model, and The steps include inputting the pre-processed weather characteristics of the target housing complex into a prediction model and obtaining prediction results, A method for predicting multi-energy loads in an integrated energy system without historical data, performed by a computer, characterized by including the following:
2. The Metas training policy specifically states: A method for predicting multi-energy loads in the absence of historical data for an integrated energy system, as described in claim 1, characterized by considering parameter updates in different source domains and fully applying the loss values and weight values of each training task in the inner layer to the update function of the outer layer.
3. The multi-energy load prediction method for an integrated energy system without historical data, as described in claim 1, characterized in that the multi-energy load prediction model comprises an encoder module and a decoder module.
4. The method for predicting multi-energy loads in the absence of historical data for an integrated energy system, according to claim 3, is characterized in that the encoder uses a CNN layer to perform feature extraction and dimensionality reduction by convolutional operations, and transmits the compressed data to LSTM to further capture long-term dependencies in the time series.
5. A method for predicting multi-energy loads in the absence of historical data for an integrated energy system, as described in claim 4, characterized by dynamically weighting the relationships between different time steps using a multi-head attention mechanism, thereby assigning 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.
6. A data acquisition and preprocessing module is configured to acquire weather characteristics of the target housing complex and historical data on air conditioning, heating, electricity, and gas for the source domain group housing complexes, and to preprocess the acquired data. A source domain data determination module is configured to perform cross-correlation and generalization capability analysis on pre-processed source domain group housing complex cooling, heating, electricity, and gas history data, and to determine appropriate source domain data. Specifically, the analysis of the intercorrelations and generalization capabilities of housing complexes is as follows: A source domain data determination module that uses the Time2vec algorithm to transform multiple load time series data from each source domain into two parts: periodic and linear, processes them using the MIC algorithm and MD algorithm respectively, constructs K-means cluster analysis to perform generalized potential stratification on the source domains, removes low-potential clusters to reduce noise interference, performs weighted fusion on MD and MIC, finds the optimal parameter range using grid search, obtains cross-correlation scores and reasonable source domain ranges for the source domain groups, and obtains the probability that a source domain can characterize a target domain using Bayesian weighted probabilities. A model building and training module is configured to construct a multi-energy load prediction model, train the model based on source domain data using the Metas training policy, adjust gradient weights based on the validation loss of the internal loop task and the source domain generalization probability, and obtain a trained prediction model. A model output module is configured to input the pre-processed weather characteristics of the target housing complex into a prediction model and obtain prediction results. A multi-energy load prediction system for integrated energy systems in the absence of historical data, characterized by including the following:
7. An electronic device comprising memory, a processor, and computer instructions stored in memory and executed on the processor, wherein the method according to any one of claims 1 to 5 is completed when the computer instructions are executed by the processor.
8. A computer-readable storage medium used for storing computer instructions, characterized in that when the computer instructions are executed by a processor, the method according to any one of claims 1 to 5 is completed.
9. A computer program product comprising a computer program, characterized in that when the computer program is executed by a processor, it realizes the completion of the method according to any one of claims 1 to 5.