Depth-generating integrated energy system source-load prediction method and system

Abstract

The present invention relates to the field of integrated energy systems technology and provides a depth generation integrated energy system source-load prediction method and system, which includes acquiring source-load data and meteorological data at several consecutive historical time points of an integrated energy system, and obtaining source-load data for future time points by a main prediction model, wherein the training step of the main prediction model includes training a discriminator as an adversarial target of the generator, based on true samples in the training set, instructing the generator to learn to generate virtual samples through adversarial training, integrating the true samples and virtual samples and inputting them into the main prediction model, and inputting the output of the main prediction model into the discriminator, using the discriminator to determine the truthfulness of the prediction results and instructing the training of the main prediction model. The generalization ability of the target model and the prediction accuracy for source-load prediction can be significantly improved.

Description

Technical Field

[0001] "Cross-reference to Related Applications" This invention claims the priority of a Chinese patent application filed with the China National Intellectual Property Administration on September 22, 2025, with the application number 202511349374.9 and the invention title "Method and System for Predicting the Source-Load of a Depth Generation Integrated Energy System", all of whose contents are incorporated into this invention by reference, constitute a part of this invention, and are used for all purposes.

[0002] This invention belongs to the technical field of integrated energy systems, and particularly relates to a method and system for predicting the source-load of a depth generation integrated energy system.

Background Art

[0003] The description of this part only provides background technical information related to this invention and does not necessarily constitute prior art.

[0004] The development of energy must balance both efficient utilization and environmental protection. As a typical paradigm of an energy system, an integrated energy system (IES) can achieve multi-functional complementarity and the integration of source-network-charge-back based on advanced cyber-physical technologies, and thus has important advantages in aspects such as improving energy utilization efficiency and promoting the development of renewable energy. Accurate source-load prediction is the first prerequisite for IES planning, operation control, and schedule setting. However, with the improvement of the penetration rate of renewable energy and the rapid popularization of loads such as electric vehicles, the source-load characteristics of IES have become more diverse and dynamic, posing great challenges to prediction.

[0005] Through the collaborative efforts of many scholars, IES prediction has gradually moved beyond the dilemma of directly applying power system prediction methods, evolving from early statistical learning to machine learning and finally to a new stage led by deep learning. In recent years, deep learning-based prediction models often use sequence neural networks as a backbone to capture time-dependent features, and are aided by convolutional neural networks, attention mechanisms, and multi-task learning (MTL) frameworks to uncover complex coupling relationships between multiple variables. Representative components include convolutional neural networks (CNNs), residual neural networks (ResNets), transformer networks, long short-term memory (LSTM) networks, timely attention mechanisms, and spatial attention mechanisms. Here, the Transformer, based on the self-attention mechanism, exhibits higher parallel computing capabilities and powerful feature representation capabilities when processing long sequence data, and has rapidly become a mainstream method in the field of prediction since its introduction. However, wind, photorenewable energy, and multi-energy loadings possess strong randomness and periodicity, and they are coupled to each other through IES multi-energy complementary mechanisms, forming a complex time-series structure with multiple layers and dynamic relationships. Transformers, on the other hand, primarily focus on long-term correlations within sequences, neglecting the description of coupling relationships and local features between different sequences. Overall, supported by multiple backbone models, researchers have obtained accurate prediction results through rational architecture design. However, much of the research still focuses on improving the network architecture itself, with little involvement in detailed examination of training mechanisms and data modeling.

[0006] As a primary prerequisite for driving model learning, the quality and scale of data directly determine the upper limit of model performance. In the IES field in particular, complex dynamic coupling relationships exist between source-load and multi-energy loads, manifesting as highly time-varying and multivariate coupling. Deep learning models can only effectively learn complex mapping features when supported by a sufficient amount of high-quality historical data. Not only data constraints, but also the setting of optimization targets are significant bottlenecks that limit the performance of predictive models. Many conventional IES prediction methods rely solely on fixed regression optimization targets such as mean squared error, L1 loss, L2 loss, and Huber loss (smooth L1 loss). Such inflexible optimization targets tend to lead to models overfitting to local modes in the historical trajectory, ignoring the modeling of the overall sequence distribution and making it difficult to adapt to the strong randomness of source-load data. [Overview of the Initiative] [Problems that the invention aims to solve]

[0007] To solve the technical challenges present in the background technologies described above, the present invention provides a depth generation integrated energy system source-load prediction method and system, and proposes a depth generation prediction framework, which consists of an auxiliary trainer and a main prediction model, and by enhancing the training process, the generalization ability of the target model and the prediction accuracy for source-load prediction can be significantly improved. [Means for solving the problem]

[0008] To achieve the above objective, the present invention employs the following technical approach.

[0009] According to a first aspect of the present invention, a depth generation integrated energy system source-load prediction method is provided, which is, To obtain source-load data and weather data at several consecutive historical time points in an integrated energy system, This includes obtaining source-load data for future times using a primary forecasting model based on several consecutive historical time points and weather data. Here, the training step of the principal predictor model includes training the discriminator to learn to generate virtual samples through adversarial training, using true samples in the training set as the adversarial target of the generator; integrating the true samples and virtual samples, inputting them into the principal predictor model, and inputting the output of the principal predictor model into the discriminator, using the discriminator to determine the truthfulness of the prediction results, and guiding the training of the principal predictor model.

[0010] Furthermore, both the generator and the main prediction model employ an encoder, and the encoder uses an embedded layer to map the input to a fixed dimension, then performs a 2-way parallel process to add the output of the embedded layer and the 2-way branch output, followed by layer normalization, learnable position coding, a multi-head attention layer and a pre-feedback layer to obtain the encoder output, and then uses residual connection and layer normalization after the multi-head attention layer and pre-feedback layer.

[0011] Furthermore, the training of the principal prediction model employs a weighted sum of weighted multitask loss and adversarial regularization loss, and the weights of the adversarial regularization loss are λ D =max(min(λ max ,λ base (1-ρ)), λ min ) and here, λ max , λ min and λ base These are the maximum adversarial loss weight, minimum adversarial loss weight, and baseline adversarial loss weight, respectively, and ρ is the similarity in the gradient direction.

number

number

number

[0012] Furthermore, the source-load data includes electrical loads, thermal loads, solar power generation, and wind power generation.

[0013] Furthermore, the weather data includes average temperature, wind speed, and irradiance.

[0014] According to a second aspect of the present invention, a depth generation integrated energy system source-load prediction system is provided, which is, A training module is configured to use true samples in a training set to train a discriminator as an adversarial target for a generator, to instruct the generator to learn to generate virtual samples through adversarial training, to integrate the true samples and virtual samples, to input the results into the main predictive model, and to input the output of the main predictive model into the discriminator, to use the discriminator to determine the truthfulness of the prediction results, and to instruct the training of the main predictive model. A data acquisition module configured to acquire source-load data and weather data at several consecutive historical time points in an integrated energy system, It includes a forecasting module configured to obtain source-load data for future times based on several consecutive historical time points and weather data, with a primary forecasting model.

[0015] Furthermore, both the generator and the main prediction model employ an encoder, and the encoder uses an embedded layer to map the input to a fixed dimension, then performs a 2-way parallel process to add the output of the embedded layer and the 2-way branch output, followed by layer normalization, learnable position coding, a multi-head attention layer and a pre-feedback layer to obtain the encoder output, and then uses residual connection and layer normalization after the multi-head attention layer and pre-feedback layer.

[0016] Furthermore, the training of the principal prediction model employs a weighted sum of weighted multitask loss and adversarial regularization loss, and the weights of the adversarial regularization loss are λD = max(min(λ max , λ base ·(1 - ρ)), λ min ), where λ max , λ min and λ base are the maximum adversarial loss weight, the minimum adversarial loss weight, and the reference adversarial loss weight respectively, ρ is the similarity in the gradient direction,

Number

Number

Number

[0017] Further, the source - load data includes an electrical load, a thermal load, solar power generation, and wind power generation.

[0018] Further, the weather data includes an average temperature, a wind speed, and an irradiance.

Advantages of the Invention

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0020] The present invention proposes a depth - generation prediction framework, which is composed of an auxiliary trainer and a main prediction model. By strengthening the training process, the generalization ability of the target model and the prediction accuracy for source - load prediction can be significantly improved.

[0021] Based on the principal regression loss function, this embodiment intends to construct an adversarial regularization loss based on a discriminator to perform a global structure and deeper level of modeling on the input data. Adversarial regularization loss allows the model to improve its ability to adapt to unknown data rather than fitting the training data more accurately. [Brief explanation of the drawing]

[0022] The accompanying drawings of the specification, which constitute part of the present invention, are for the purpose of providing a further understanding of the present invention, and the exemplary embodiments and descriptions thereof are for the purpose of interpreting the present invention and do not constitute an unreasonable limitation of the present invention. [Figure 1] This is a flowchart of a depth generation integrated energy system source-load prediction method according to Embodiment 1 of the present invention. [Figure 2] This is a diagram of the encoder structure according to one embodiment of the present invention. [Figure 3] This is a comparison diagram of the electrical load prediction results for one embodiment of the present invention. [Figure 4] This is a comparison diagram of the heat load prediction results for one embodiment of the present invention. [Figure 5] This is a comparison diagram of the solar power generation prediction results for one embodiment of the present invention. [Figure 6] This is a comparison diagram of wind power generation prediction results for one embodiment of the present invention. [Modes for carrying out the invention]

[0023] To further clarify the purpose, technical proposal, and advantages of the embodiments of the present invention, the following clearly and completely describes the technical proposal in the embodiments of the present invention, linking it to the accompanying drawings.

[0024] The following detailed descriptions are illustrative and intended to further illustrate the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those generally understood by those skilled in the art.

[0025] Example 1 This embodiment provides a depth generation integrated energy system source-load prediction method.

[0026] The depth generation integrated energy system source-load prediction method according to this embodiment solves the problems of data scarcity and localization of optimization targets that current deep learning prediction models face during the training phase through a depth generation prediction framework.

[0027] The depth-generating integrated energy system source-load prediction method according to this embodiment is inspired by the successful generation and regularization of Generative Adversarial Networks (GANs). To address the shortcomings of data scarcity and localization of optimization targets that exist during predictive model training, we propose a Deep Generative Prediction Framework (DGP). This framework consists of an auxiliary trainer and a main predictive model, which can be flexibly integrated into any conventional predictive model, and by enhancing the training process, the predictive accuracy and generalization ability of the target model can be significantly improved.

[0028] GANs are generative models based on game theory principles, consisting of generators and discriminants. They generate high-fidelity samples that match the true data by improving the model's ability to depict data distribution information through adversarial optimization. Inspired by this, this embodiment introduces an adversarial generative mechanism to time series prediction tasks, constructs an auxiliary trainer, and extends the training process from two perspectives: data augmentation and regularization of the optimization target. To address the instability and vanishing gradient problem often seen during conventional GAN ​​training, this embodiment employs an improved adversarial generative network (Wasserstein GAN with Gradient Penalty, WGAN-GP), using Earth-Mover (bulldozer) distance as an evaluation metric and employing a gradient penalty term instead of the original weight trimming mechanism.

[0029] As shown in Figure 1, the DGP framework proposed in this embodiment is a training extension structure based on generative-adversarial thinking, consisting of an auxiliary trainer and a primary predictor model, where the primary predictor model can be arbitrarily replaced with other predictor models. In this framework, the generator expands the diversity of the training data by generating pseudo-samples that are more useful for source-load trend modeling, based on GAN adversarial training and further linked with gradient information of the primary predictor model for reverse induction fine-tuning. The discriminator regularizes the primary predictor model at the array level and provides discriminative feedback at the global distribution level, guiding it to model the global structure of the source-load trend. As the core of the framework, the primary predictor model completes future trend modeling through cooperative adversarial interaction with the auxiliary trainer. In this embodiment, in order to better adapt the large number of samples generated by the auxiliary trainer, the primary predictor model is further proposed, forming a shared representation backbone based on the superposition of multi-layer Cencoders and linked with a multi-output MTL (multi-task learning) architecture to achieve cooperative modeling and accurate prediction for different variable distributions.

[0030] The depth generation integrated energy system source-load prediction method according to this embodiment specifically includes the following steps:

[0031] Step 1: Build a generator based on Cencoder and generate virtual training samples that combine true statistical characteristics with unknown change modes—especially edge scenes that conform to system operating rules, such as when the landscape output changes abruptly and the load fluctuates extremely.

[0032] Step 101, the Transformer excels at capturing long-range time dependencies through its Multi-Head Self-Attention (MHSA) mechanism, possessing powerful global modeling capabilities and the advantages of parallel computing. To further enhance its feature extraction capabilities, a one-dimensional convolutional neural network (1D convolutional neural network, Conv1D) is introduced as a complementary module, proposing the encoder-Cencoder. In DGP, the generator can synthesize large amounts of data, enabling the construction of deeper levels of networks to extend the model's representational capabilities. However, deep models are prone to the vanishing gradient and training degeneration problems. Therefore, a multi-stage residual connection mechanism is introduced into the Cencoder to extend gradient fluidity and improve training efficiency. The overall structure of the Cencoder is shown in Figure 2, and includes a convolutional module, a learnable position coding module, and an Encoder (coding) module.

[0033] (1) Convolutional module.

[0034] First, the embedding layer is the input array

number

number

[0035] After that, X embed This involves a two-way parallel processing, Branch 1 extracts join features between different sequences using Conv1D and MaxPool1D, stabilizes the distribution using Norm, and the output is: H conv1 =Norm(MaxPool1D(Conv1D(X embed))) (2) can be expressed as, Note X embed This transforms the dimensions to B×C×L before inputting them into the convolutional feature extraction module, where C is the number of channels, and restores them after the convolution operation is complete. Branch 2 directly applies one-dimensional convolution to the embedding representation, preserving local dependencies and fine-grained dynamic change information between adjacent time steps without pooling, and the output is: H conv2 =Conv1D(X embed ) can be expressed in (3), Finally, to preserve the original features of the data, the original embedding and the binary output are added together and then the convolutional module output is obtained via Layer Normalization (LayerNorm).

number

[0036] (2) Learnable position coding.

[0037] Considering the insensitivity of the self-attention mechanism to the order of arrangement, X conv It is necessary to inject positional information into it. However, conventional position coding is mainly used to express the priority relationship of positions in an array, X conv This not only includes array data, but also other high-dimensional local feature representations, resulting in more complex spatial-temporal distribution characteristics. Therefore, we employ learnable position coding, improving the modeling capability for multiscale features by having the model automatically adjust the representation of each position based on training. The learnable position coding output is, X pos =X conv It can be expressed as +P(6), Here,

number

[0038] (3) Encoder module.

[0039] X after joining location coding pos The data is transmitted to the Encoder, where long-term dependent feature extraction is performed. Each layer of the Encoder includes two sublayers: MHSA (Multi-Head Attention Layer) and a feedforward network (pre-feedback layer), and residual connection and layer normalization are used after each sublayer.

[0040] The core of MHSA is the scaled inner product attention formula.

number

number

number

[0041] The i-th self-attention output head i =Attention(Q i ,K i ,V i If we define it as follows, the MSHA output will be: MSHA(Q,K,V)=Concat(head1,...,head h )W o (9) Here, h is the number of heads,

number

[0042] The MSHA sublayer output is obtained by residual connection and layer normalization. X MSHA =LayerNorm(X pos +MHSA(Q,K,V)) (10) can be obtained, After that, X MSHA It enters the feedforward network, is typically a two-layer all-connection layer, and is the output of the feedforward network. X FFN =ReLU(ReLU(X MSHA We obtain (11) (W1+b1)W2+b2), Here, ReLU is a normalized linear function,

number

number

number

number

[0043] Finally, after going through residual connection and layer normalization again, we get the Cencoder output. X C =LayNorm(X MSHA +X FFN ) (12) can be obtained, Step 102, Generator.

[0044] Based on Cencoder construction, the generator effectively models global time-series dependencies and local coupling features in the source-load sequence, thereby generating virtual training samples that possess true statistical properties and unknown modes of change—especially edge scenes that conform to system operating rules, such as when the landscape output changes abruptly and the load fluctuates drastically. To ensure data controllability, this embodiment employs a conditional generation policy, inputting historical data as generation conditions along with noise vectors into the generator network.

[0045] The generation process is,

number

number

number

number

number

[0046] The generator's target is to generate samples that can "fake" the discriminator, i.e., to minimize the score of the generated samples, and the loss function is:

number

number

number

[0047] Since the prediction task belongs to supervised learning, the samples generated by the generator consist of predicted input and output tags.

[0048] Step 2: Construct a discriminator based on a one-dimensional convolutional network. The discriminator is used as an adversarial target for the generator, and adversarial training guides the generator to learn to generate high-quality virtual samples, thereby improving its own ability to discriminate between true and generated samples. Together, the generator and discriminator constitute an auxiliary trainer.

[0049] The discriminator, acting as an adversarial target for the generator, not only improves the truthfulness of the generated samples through a game process, but also, as an auxiliary structure, provides regularization feedback to the main predictive model regarding the sample distribution. It should be noted that updating the discriminator parameters depends only on the adversarial training phase with the generator and remains frozen during the main model training, avoiding the additional introduction of gradient interference.

[0050] In the discriminator, the input first undergoes a combination of processing in one-dimensional convolutional, pooling, and normalization layers. Next, a flattening layer converts multi-dimensional features into one-dimensional feature vectors. Finally, after passing through a Linear layer and a Sigmoid layer, the Sigmoid (S-type) function maps the output value between 0 and 1, thereby determining whether the input data is true (real data) or false (false data generated by the generator).

[0051] The discrimination process is,

number

number

[0052] The target of the discriminator is to distinguish between true data and input data, and its adversarial loss function adopts the WGAN-GP form, with the definition being:

number

number

number

[0053] Step 3: Implement the primary predictive model. As the core of the framework, the primary predictive model completes future trend modeling by extracting mapping features from the input data through collaborative adversarial interaction with the auxiliary trainer. The primary predictive model can be optionally replaced with other conventional predictive models.

[0054] Thanks to the large number of pseudo-samples generated by the generator, the primary predictive model can employ a deeper network structure and extract more expressive deep features without overfitting problems, by having stronger data support during training. Specifically, the input array is first processed by a multi-layered Cencoder to extract deep-order shared representations, forming a feature representation space rich in meaning. The shared representation from the final time step is then used to construct global shared features, which serve as the common basis for multiple task outputs.

[0055] Enter

number

number

[0056] The definition of the principal loss function is the weighted multitask loss,

number

number

number

number

[0057] Based on the principal regression loss function, this embodiment uses a discriminant to determine the adversarial regularization loss L D The aim is to build a model that provides a global structure and deeper level of modeling for the input data. Adversarial regularization loss allows the model to improve its ability to adapt to unknown data rather than fitting the training data more accurately.

[0058] The total loss function is,

number

number

[0059] Step 4: Integrate the virtual samples generated by the generator with the true historical data to enhance the data diversity of the training set and input it into the main predictive model for training.

[0060] Step 5: The input first goes through a multitask learning layer of the main prediction model to learn shared parameters between different subtasks and achieve multitask collaborative optimization. Then, it enters the feature interpretation layer, where multiple prediction heads map the shared features to their respective target output spaces to achieve differentiated feature representation.

[0061] Step 6: The output information of the primary predictive model is fed back to the generator, which induces secondary fine-tuning to progressively optimize its generation policy. This process ensures that the output samples are more useful for learning the primary task, thereby achieving collaborative optimization of sample generation and feature modeling.

[0062] The second-order fine-tuning loss function is:

number

[0063] Simultaneously, the output of the primary prediction model is input to a classifier, which is used to determine the truthfulness of the prediction results and to guide the training of the primary prediction model.

[0064] This embodiment validates the proposed DGP in an IES project within a housing complex, including time series of electrical load, thermal load, solar power generation, wind power generation, mean temperature, wind speed, and irradiance. The dataset includes 1825 days of hourly source-load and meteorological measurement data, with a time range from January 1, 2015 to December 31, 2019. To comprehensively evaluate the DGP's performance in the source-load forecasting task, four evaluation metrics are employed: mean root square error (RMSE), mean absolute percent error (MAPE), and coefficient of determination (R²). These metrics cover multiple dimensions of error range and fitting ability, allowing for an integrated assessment of model performance from different perspectives.

[0065] To evaluate the effectiveness of the proposed DGP in 24 and 168h continuous prediction tasks, nine advanced models are compared against a baseline. These baseline models include not only conventional instrument learning methods such as SVR (Support Vector Regression), but also the most advanced improved Transformer-based prediction models, such as Informer, Preformer, and Autoformer. Simultaneously, classical prediction models based on two RNNs (Circular Neural Networks), LSTM (Long Short-Term Memory Network) and TCN (Time-Series Convolutional Network), are considered. Furthermore, general prediction models based on convolutional structures, such as CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory Network) and GCN (Graphical Attention Circular Neural Network) and T-GCN (Time-Graphical Convolutional Network), are introduced. Note that the above structures are only the backbone mechanisms; in reality, the baseline models also include auxiliary structures such as attention mechanisms and multi-task learning. The hyperparameters of the baseline models are determined by the mesh search method.

[0066] Table 1, Results of comparative tests [Table 1]

[0067] As can be seen from Table 1, the proposed DGP exhibits a significant performance advantage in 24-hour and 168-hour continuous forecasting tasks for electrical and thermal loads, and wind and optical power. Taking electrical loads as an example, in the 24-hour continuous forecasting task, the MAPE and RMSE of DGP were reduced by 36.4% and 37.6%, respectively, compared to the best-performing baseline model, Preformer. R2 was significantly better than Autoformer's 0.9287. In the 168-hour forecasting task, the MAPE and RMSE of DGP were reduced by 34.55% and 55.86%, respectively, compared to the best-performing baseline model, Preformer, and R2 was significantly better than Preformer's 0.8423. More importantly, while the baseline models showed a rapid increase in error in long-duration continuous forecasting tasks, DGP demonstrated remarkable stability and anti-error diffusion capability, with the smallest difference in accuracy between the 24-hour and 168-hour continuous forecasting tasks. The higher accuracy of CNN-LSTM compared to conventional LSTM demonstrates that adding a convolutional module to a cyclic structure effectively improves the model's ability to extract local features, which is one of the inspirations for the Cencoder design in this embodiment. In all baseline models, the Transformer series models outperform conventional predictive models overall, showing a significant advantage in source-load time series tasks with complex couplings. To illustrate the prediction results more intuitively, some representative models were selected and visualized with true data (Actual), as shown in Figures 3, 4, 5, and 6.

[0068] Example 2 This embodiment provides a depth generation integrated energy system source-load prediction system, which specifically, A training module is configured to use true samples in a training set to train a discriminator as an adversarial target for a generator, to instruct the generator to learn to generate virtual samples through adversarial training, to integrate the true samples and virtual samples, to input the results into the main predictive model, and to input the output of the main predictive model into the discriminator, to use the discriminator to determine the truthfulness of the prediction results, and to instruct the training of the main predictive model. A data acquisition module configured to acquire source-load data and weather data at several consecutive historical time points in an integrated energy system, It includes a forecasting module configured to obtain source-load data for future times based on several consecutive historical time points and weather data, with a primary forecasting model.

[0069] Furthermore, both the generator and the main prediction model employ an encoder, and the encoder uses an embedded layer to map the input to a fixed dimension, then performs a 2-way parallel process to add the output of the embedded layer and the 2-way branch output, followed by layer normalization, learnable position coding, a multi-head attention layer and a pre-feedback layer to obtain the encoder output, and then uses residual connection and layer normalization after the multi-head attention layer and pre-feedback layer.

[0070] Furthermore, the primary predictive model training employs a weighted sum of weighted multitask loss and adversarial regularization loss, where the weights of the adversarial regularization loss are λ. D =max(min(λ max ,λ base (1-ρ)), λ min ) and here, λ max , λ min and λ base These are the maximum adversarial loss weight, minimum adversarial loss weight, and baseline adversarial loss weight, respectively, and ρ is the similarity in the gradient direction.

number

number

number

[0071] Furthermore, the source-load data includes electrical loads, thermal loads, solar power generation, and wind power generation.

[0072] Furthermore, the weather data includes average temperature, wind speed, and irradiance.

[0073] Each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and the specific implementation process is the same; therefore, no further explanation is provided here.

[0074] The foregoing describes only preferred embodiments of the present invention and is not intended to limit it. Those skilled in the art will know that various modifications and changes are possible to the present invention. Any modifications, equivalent substitutions, or improvements must be within the scope of protection of the present invention, within the spirit and principles of the invention.

[0075] The foregoing describes only preferred embodiments of the present invention and is not intended to limit it. Those skilled in the art will know that various modifications and changes are possible to the present invention. Any modifications, equivalent substitutions, or improvements must be within the scope of protection of the present invention, within the spirit and principles of the invention.

Claims

1. A depth generation integrated energy system source-load prediction method, To acquire source-load data and weather data for an integrated energy system at several consecutive historical time points, This includes obtaining source-load data for future times using a primary forecasting model based on several consecutive historical time points and weather data. Here, the training step of the principal prediction model includes, using true samples in the training set, training the discriminator to learn to generate virtual samples through adversarial training, treating the discriminator as an adversarial target for the generator, integrating the true samples and virtual samples, inputting them into the principal prediction model, inputting the output of the principal prediction model into the discriminator, using the discriminator to determine the truthfulness of the prediction results, and guiding the training of the principal prediction model. The aforementioned primary predictive model training employs a weighted sum of weighted multitask loss and adversarial regularization loss, and the weights of the adversarial regularization loss are λ Ｄ = max(min(λ) ｍａｘ ,λ ｂａｓｅ (1-ρ),λ ｍｉｎ ) and here, λ ｍａｘ , λ ｍｉｎ and λ ｂａｓｅ These are the maximum adversarial loss weight, minimum adversarial loss weight, and baseline adversarial loss weight, respectively, and ρ is the similarity in the gradient direction. [Math 1] And, [Math 2] This is a weighted multitasking loss, [Math 3] A depth generation integrated energy system source-load prediction method characterized in that the loss is adversarial regularization loss.

2. The depth generation integrated energy system source-load prediction method according to claim 1, characterized in that both the generator and the main prediction model employ an encoder, and the encoder employs an embedded layer to map the input to a fixed dimension, then performs a two-way parallel process to add the output of the embedded layer and the two-way branch output, and then sequentially passes through layer normalization, learnable position coding, a multi-head attention layer and a pre-feedback layer to obtain the encoder output, and then uses residual connection and layer normalization after the multi-head attention layer and the pre-feedback layer.

3. The depth generation integrated energy system source-load prediction method according to claim 1, characterized in that the source-load data includes an electrical load, a thermal load, solar power generation, and wind power generation.

4. The depth generation integrated energy system source-load prediction method according to claim 1, characterized in that the weather data includes average temperature, wind speed, and irradiance.

5. Depth generation integrated energy system source-load prediction system, A training module configured to use true samples in a training set to train a discriminator as an adversarial target for a generator, to instruct the generator to learn to generate virtual samples through adversarial training, to integrate the true samples and virtual samples and input them into a principal predictor model, and to input the output of the principal predictor model into the discriminator, to use the discriminator to determine the truthfulness of the prediction results, and to instruct the training of the principal predictor model, The main prediction model training adopts a weighted sum of the weighted multitask loss and the adversarial regularization loss, and the weight of the adversarial regularization loss is λ Ｄ = max(min(λ ｍａｘ , λ ｂａｓｅ ·(1 - ρ)), λ ｍｉｎ ), where λ ｍａｘ , λ ｍｉｎ and λ ｂａｓｅ are the maximum adversarial loss weight, the minimum adversarial loss weight and the reference adversarial loss weight respectively, and ρ is the similarity in the gradient direction. [Math 4] And, [Math 5] This is a weighted multitasking loss, [Math 6] The training module is the adversarial regularization loss, A data acquisition module configured to acquire source-load data and weather data at several consecutive historical time points in an integrated energy system, A depth-generating integrated energy system source-load forecasting system characterized by including a forecasting module configured to obtain source-load data for future times by a main forecasting model based on source-load data and meteorological data for several consecutive historical times.

6. The depth generation integrated energy system source-load prediction system according to claim 5, characterized in that both the generator and the main prediction model employ an encoder, and the encoder employs an embedded layer to map the input to a fixed dimension, then performs a two-way parallel process to add the output of the embedded layer and the two-way branch output, and then sequentially passes through layer normalization, learnable position coding, a multi-head attention layer and a pre-feedback layer to obtain the encoder output, and then uses residual connection and layer normalization after the multi-head attention layer and the pre-feedback layer.

7. The depth generation integrated energy system source-load prediction system according to claim 5, characterized in that the source-load data includes an electrical load, a thermal load, solar power generation, and wind power generation.

8. The depth generation integrated energy system source-load prediction system according to claim 5, characterized in that the weather data includes average temperature, wind speed, and irradiance.

Images