Full-modal hybrid expert model energy prediction method based on hierarchical routing

Abstract

The invention relates to the technical field of energy prediction, and discloses a hierarchical routing-based full-modal hybrid expert model energy prediction method, which comprises the following steps: S1, constructing a hierarchical expert network structure which comprises a plurality of expert layers, and each expert layer is configured to process feature data of a specific modal, wherein the expert layer is divided into a meteorological processing layer, an electricity price processing layer, a load processing layer and an illumination processing layer based on modal types; and S2, receiving multi-modal characteristic data, wherein the multi-modal characteristic data comprises meteorological data, electricity price data, load data and illumination data. A layered expert network structure and a dynamic routing distribution mechanism are constructed, a special expert layer is adopted to perform feature extraction and local prediction, and a dynamic distribution strategy of a routing module is combined, so that the compatibility of a prediction model to high-dimensional heterogeneous data is improved, the calculation complexity of model training and deployment is reduced, and the prediction efficiency is improved. And the real-time performance and resource constraint requirements in a virtual power plant scene are met.

Description

Technical Field

[0001] The present invention relates to the technical field of energy prediction, and in particular to a full-modal hybrid expert model energy prediction method based on hierarchical routing. Background Art

[0002] In the energy management scenario of a virtual power plant, accurate prediction of load demand, power generation, and carbon emission-related indicators is the key to achieving energy-carbon coordinated optimization. Energy forecasting often relies on a single model or a simple multi-source data fusion method, such as traditional time series analysis or general machine learning models, aiming to improve the reliability of forecast results through data-driven strategies. However, with the increasing complexity of energy systems, especially the diversity and heterogeneity of multimodal data, the limitations of traditional methods have gradually become apparent.

[0003] When processing multimodal data, current energy forecasting methods often adopt a unified model or shallow feature fusion strategy, lacking the ability to adapt to different modal features in a targeted manner. Characteristics such as the time dependence of meteorological data, the market volatility of electricity price data, and the spatiotemporal correlation of load data are not effectively distinguished, resulting in insufficient compatibility of the model with high-dimensional heterogeneous data and susceptibility to data coupling interference during the forecasting process. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a full-modal hybrid expert model energy prediction method based on hierarchical routing, which solves the problem of lack of targeted adaptation capability for heterogeneous features, resulting in insufficient model compatibility and high computational complexity, making it difficult to support real-time and efficient prediction needs in virtual power plant scenarios.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a full-modal hybrid expert model energy prediction method based on hierarchical routing, comprising the following steps: S1. Constructing a hierarchical expert network structure, wherein the hierarchical expert network structure includes multiple expert layers, each expert layer is configured to process feature data of a specific modality, wherein the expert layers are divided into a meteorological processing layer, an electricity price processing layer, a load processing layer, and an illumination processing layer based on the modality type; S2. Receive multimodal feature data, including meteorological data, electricity price data, load data, and light data, and preprocess the data to generate a standardized time series data stream; S3. Input the standardized time series data stream into a routing module. The routing module dynamically allocates each modal data to a corresponding expert layer based on a feature similarity measure between a preset modal label and the input data. The output of the routing module is a mapping relationship between each modal data and its target expert layer. S4. In each expert layer, feature extraction and local prediction are performed on the assigned modal data to generate expert layer local prediction results, wherein the local prediction results include load prediction values, power generation prediction values, and carbon emission related indicators; S5. Receive the local prediction results of all expert layers through the decision fusion module, dynamically adjust the weights based on the historical prediction accuracy of each expert layer, use the attention mechanism to perform weighted fusion on the local prediction results, and output the final energy prediction result; S6. Feedback the energy forecast results to the virtual power plant dispatching system for real-time adjustment of power generation plans, load distribution or carbon emission optimization strategies.

[0006] Preferably, in step S1, the expert layer adopts a heterogeneous model structure, including at least one of the following combinations: the meteorological processing layer adopts a long short-term memory neural network, the electricity price processing layer adopts a gradient boosting decision tree, the load processing layer adopts a convolutional neural network, and the illumination processing layer adopts support vector regression.

[0007] Preferably, in step S3, the dynamic allocation logic of the routing module includes: Calculate the cosine similarity between the feature vector of the input data and the preset modal labels of each expert layer; Select the expert layer with the highest similarity as the target routing path; If the similarity is lower than the preset threshold, the exception handling mechanism is triggered and the data is routed to the default expert layer for cross-modal joint analysis.

[0008] Preferably, in step S5, the attention mechanism includes the following steps: Generate initial weights based on the historical prediction errors of each expert layer; Dynamically scoring the confidence of the local prediction results through a multi-layer perceptron to correct the initial weights; The local prediction results are weighted and summed according to the corrected weights to generate the final prediction result.

[0009] Preferably, in step S5, the decision fusion module includes the following steps: Based on the local prediction results of each expert layer generated by S4, the historical prediction error records of each expert layer are extracted to generate the initial weight distribution table; Associating the data mapping relationship output by the routing module in S3 with the initial weight distribution table, dynamically scoring the confidence of the local prediction results through a multi-layer perceptron, and correcting the initial weights; The local prediction results of S4 are weighted and fused according to the revised weights to generate the final energy prediction results; Receive real-time operational data from S6's virtual power plant dispatching system, including actual load values, power generation values, and carbon emission monitoring results; Performing an error comparison between the operating data and the final energy forecast result to generate a model deviation index; Based on the model deviation index, the model parameters of each expert layer in step S1 are updated through the incremental learning algorithm, and the similarity threshold of the routing module in S3 and the weight allocation strategy of the decision fusion module in S5 are adjusted synchronously.

[0010] Preferably, in step S2, the preprocessing includes: Normalize the temperature, humidity and wind speed in meteorological data; Perform time series alignment and missing value interpolation on electricity price data; The load data and light data are subjected to outlier removal and smoothing filtering respectively.

[0011] Preferably, in step S4, the local prediction results of the expert layer are further used to generate an intermediate decision report, which includes: Feature extraction paths and processing logic for each expert layer; Dynamic allocation records of routing modules; The contribution ratio of local prediction results to the final prediction is used to verify the model transparency.

[0012] Preferably, the final energy forecast result includes at least one of the following: Load demand curve for the next 24 hours; Probability distribution of photovoltaic power generation; Carbon emission sensitivity matrix, used to evaluate the carbon economy of different dispatch strategies.

[0013] Preferably, the training of the expert layer adopts a phased joint optimization strategy: In the first stage, each expert layer is trained independently and the model parameters are optimized using the historical data of the corresponding modality; In the second stage, the expert layer parameters are fixed, and the routing module and decision fusion module are trained end-to-end to minimize the overall prediction error; The third stage jointly fine-tunes all modules to improve the generalization ability of the model.

[0014] Preferably, the method is deployed in an edge-cloud collaborative architecture: Edge devices perform data preprocessing, routing assignment, and expert-level local computation; The cloud server receives the local prediction results of each edge device and performs decision fusion and global optimization; The cloud sends the optimized scheduling instructions to the edge device to form a closed-loop control link.

[0015] The present invention provides a full-modal hybrid expert model energy prediction method based on hierarchical routing. It has the following beneficial effects: 1. The present invention realizes the targeted processing of multimodal energy data and the adaptation of heterogeneous models by constructing a hierarchical expert network structure and a dynamic routing distribution mechanism. A dedicated expert layer is used for feature extraction and local prediction based on the different modal characteristics of meteorology, electricity prices, load and illumination. Combined with the dynamic allocation strategy of the routing module, the coupling interference of multi-source data is avoided, and the multi-objective prediction problem of complex energy systems is decomposed into multiple subtasks for collaborative solution, thereby improving the compatibility of the prediction model with high-dimensional heterogeneous data and reducing the computational complexity of model training and deployment, meeting the real-time and resource constraint requirements in the virtual power plant scenario.

[0016] 2. The present invention realizes dynamic weight adjustment and online parameter optimization of the prediction model by introducing the attention mechanism and incremental learning algorithm. Based on historical prediction errors and real-time feedback data, the fusion weights of each expert layer are dynamically corrected. The model parameters and routing thresholds are iteratively updated in combination with incremental learning, so that the model can adapt to changes in the operating environment of the energy system, thereby realizing the continuous evolution capability of the prediction strategy. At the same time, through the explicit association of model deviation indicators with intermediate decision reports, the interpretability of system decisions is enhanced, providing a transparent analysis basis for dispatchers.

[0017] 3. The present invention realizes efficient utilization of distributed computing resources and enhanced model generalization capabilities through edge-cloud collaborative architecture and phased joint training strategy, performs data preprocessing, routing distribution and lightweight expert layer calculations at the edge, and centrally performs global fusion and strategy optimization on the cloud. Combined with phased training to balance model accuracy and convergence efficiency, high-density computing tasks are offloaded to the cloud. The edge only needs to process low-latency demand tasks, thereby achieving low-latency response and high scalability support for large-scale virtual power plant systems, while reducing dependence on centralized computing resources and adapting to the distributed deployment characteristics of the energy Internet of Things. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a step diagram of the energy prediction method of the full-modal hybrid expert model based on hierarchical routing of the present invention; Figure 2 Schematic diagram of the energy forecasting method of the present invention based on the full-modal hybrid expert model of hierarchical routing, in which the expert layer adopts a heterogeneous model structure combination; Figure 3 A schematic diagram of the dynamic allocation logic of the routing module of the energy prediction method based on the full-modal hybrid expert model of hierarchical routing of the present invention; Figure 4 This is a schematic diagram of the attention mechanism flow of the energy prediction method of the full-modal hybrid expert model based on hierarchical routing of the present invention; Figure 5 Schematic diagram of the final energy prediction result in the decision fusion module of the energy prediction method of the full-modal hybrid expert model based on hierarchical routing of the present invention; Figure 6 Schematic diagram of the feedback of the power plant dispatching system in the decision fusion module of the energy prediction method based on the full-modal hybrid expert model of hierarchical routing of the present invention; Figure 7 Schematic diagram of an intermediate decision report of the energy forecasting method of the full-modal hybrid expert model based on hierarchical routing of the present invention; Figure 8 Schematic diagram of the final energy prediction result of the energy prediction method of the full-modal hybrid expert model based on hierarchical routing of the present invention; Figure 9 Schematic diagram of the optimization strategy of the energy prediction method of the full-modal hybrid expert model based on hierarchical routing of the present invention; Figure 10 This is a schematic diagram of the edge-cloud collaborative architecture of the full-modal hybrid expert model energy prediction method based on hierarchical routing in the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solution of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] Please see the attached Figure 1 -Attached Figure 10 The embodiment of the present invention provides a method for energy prediction based on a full-modal hybrid expert model using hierarchical routing, comprising the following steps: S1. Construct a hierarchical expert network structure, which includes multiple expert layers. Each expert layer is configured to process feature data of a specific modality. The expert layers are divided into a meteorological processing layer, an electricity price processing layer, a load processing layer, and an illumination processing layer based on the modality type. By dividing the model into specialized expert layers for meteorological, electricity price, load, and light modalities, and designing heterogeneous models for different data characteristics, the complex multimodal prediction task is decomposed into specialized subtasks. This avoids the problem of insufficient generalization ability of a single model for different data modalities, and provides a model foundation for subsequent dynamic routing and hierarchical decision-making, enhancing the interpretability of the prediction task. S2. Receive multimodal feature data, including meteorological data, electricity price data, load data, and light data, and preprocess the data to generate a standardized time series data stream; Normalize, time-series align, and process outliers for multi-source heterogeneous data, eliminate data scale differences and noise interference, generate standardized time-series data streams, unify data input formats, reduce model deviations caused by data quality differences, provide highly consistent input for routing module distribution and expert layer feature extraction, and reduce subsequent computational complexity; S3. Input the standardized time series data stream into the routing module. The routing module dynamically assigns each modal data to the corresponding expert layer based on the feature similarity measurement between the preset modal label and the input data. The output of the routing module is the mapping relationship between each modal data and its target expert layer. Based on cosine similarity, the system matches input data with the modal labels preset by the expert layer to achieve targeted allocation of multimodal data. If the data features are ambiguous, they are routed to the integrated model for cross-modal joint analysis. This solves the problem of heterogeneous data coupling interference and ensures that each expert layer focuses on its dominant modality. At the same time, exception handling enhances system fault tolerance and avoids prediction failures caused by routing errors. S4. In each expert layer, feature extraction and local prediction are performed on the assigned modal data to generate expert layer local prediction results, where the local prediction results include load prediction values, power generation prediction values, and carbon emission related indicators; Based on the characteristics of their models, each expert layer uses CNN to capture the spatial correlation of loads and SVR to fit the nonlinear relationship of illumination. They perform feature extraction and local prediction to generate key indicators of load and power generation. This specialized model improves the accuracy of single-modal predictions, and local prediction results provide a fine-grained decision-making basis for subsequent fusion. Intermediate reports also support transparent model verification. S5. Receive the local prediction results of all expert layers through the decision fusion module, dynamically adjust the weights based on the historical prediction accuracy of each expert layer, use the attention mechanism to perform weighted fusion on the local prediction results, and output the final energy prediction result; The attention mechanism is used to dynamically adjust the fusion weights of each expert layer. Based on historical errors and real-time confidence scores, the weighted optimal fusion of prediction results is achieved. Combined with the incremental learning algorithm, the scheduling feedback data is used to update the model parameters and routing thresholds online, forming a closed loop of "prediction-feedback-tuning". This improves the model's adaptability to dynamic environments, reduces the impact of high-error expert layers through dynamic weight adjustment, and enhances overall robustness.

[0021] S6. Feedback the energy forecast results to the virtual power plant dispatching system for real-time adjustment of power generation plans, load distribution or carbon emission optimization strategies.

[0022] The final prediction results, load curve, and carbon sensitivity matrix are input into the scheduling system to guide the adjustment of power generation plans and carbon emission optimization, open up the link between prediction and execution, and enable the model output to directly act on the physical system, forming a closed-loop control of "data-driven decision-making-strategy execution-effect feedback", supporting the real-time implementation of the virtual power plant's energy-carbon synergy goals.

[0023] In step S1, the expert layer adopts a heterogeneous model structure, including at least one of the following combinations: the meteorological processing layer adopts a long short-term memory neural network, the electricity price processing layer adopts a gradient boosting decision tree, the load processing layer adopts a convolutional neural network, and the illumination processing layer adopts support vector regression.

[0024] Heterogeneous models are customized to adapt to the inherent characteristics of different modal data such as weather, electricity prices, load, and sunlight. The meteorological processing layer uses LSTM, which uses a long short-term memory network to capture the temporal dependencies and periodic changes of temperature and wind speed in meteorological data. The electricity price processing layer uses GBDT, which fits the nonlinear fluctuation of electricity price data and the market game relationship through gradient boosting decision tree; The load processing layer uses CNN to extract the spatial correlation features of load data (such as regional power consumption patterns); The illumination processing layer uses SVR and support vector regression to handle the sparsity and nonlinear mapping of illumination data; The heterogeneous model combination fully leverages the domain advantages of each algorithm, avoids the incompatibility of a single model with different modal data, significantly improves the single-modal prediction accuracy, and provides high-quality local results for subsequent fusion.

[0025] In step S3, the dynamic allocation logic of the routing module includes: Calculate the cosine similarity between the feature vector of the input data and the preset modal labels of each expert layer; Select the expert layer with the highest similarity as the target routing path; If the similarity is lower than the preset threshold, the exception handling mechanism is triggered and the data is routed to the default expert layer for cross-modal joint analysis.

[0026] Cosine similarity ensures the scientific nature of routing decisions by quantifying the degree of match between input data and preset modal labels. For example, illumination data is preferentially assigned to the illumination processing layer. In the exception handling mechanism, when data feature ambiguity occurs in the coupling of meteorological and lighting data, it is routed to the default model integrated by all expert layers for cross-modal joint analysis to avoid prediction failure due to routing errors. Through dynamic distribution and fault-tolerant mechanisms, the interference of multimodal data coupling is reduced, and the system's adaptability to data anomalies caused by extreme weather in complex scenarios is enhanced.

[0027] In step S5, the decision fusion module includes the following steps: Based on the local prediction results of each expert layer generated by S4, the historical prediction error records of each expert layer are extracted to generate an initial weight distribution table; The data mapping relationship output by the routing module in S3 is associated with the initial weight distribution table, and the confidence of the local prediction results is dynamically scored through the multi-layer perceptron to correct the initial weights; The local prediction results of S4 are weighted and fused according to the revised weights to generate the final energy prediction results; Receive real-time operational data from S6's virtual power plant dispatching system, including actual load values, power generation values, and carbon emission monitoring results; Compare the error between the operating data and the final energy forecast results to generate the model deviation index; Based on the model deviation index, the model parameters of each expert layer in step S1 are updated through the incremental learning algorithm, and the similarity threshold of the routing module in S3 and the weight distribution strategy of the decision fusion module in S5 are adjusted synchronously.

[0028] The generation of initial weights is based on historical error records and the error rate of the load forecasting layer over the past week, thereby assigning initial weights to prevent low-precision expert layers from dominating decision-making; The confidence dynamic scoring uses a multi-layer perceptron combined with the real-time electricity price fluctuation data characteristics to evaluate the reliability of the prediction results of each expert layer and dynamically adjust the weights; Weighted fusion and closed-loop feedback apply the corrected weights to local prediction results, and use scheduling feedback data to continuously optimize the weight allocation strategy, realizing dual weight control of "static experience + dynamic evaluation", balancing the impact of historical performance and real-time environmental changes, and improving the robustness of the fusion results.

[0029] In step S5, the decision fusion module performs the following steps: Based on the local prediction results of the expert layer generated in step S4, extract the prediction error history of each expert layer and generate an initial weight distribution table; Associating the data mapping relationship output by the routing module in step S3 with the initial weight distribution table, dynamically scoring the confidence of the local prediction results through the multi-layer perceptron, and correcting the initial weights; Perform weighted fusion on the local prediction results of step S4 according to the corrected weights to generate the final prediction result; Receive operational feedback data from the virtual power plant dispatching system in real time, compare the operational feedback data with the final prediction results, and generate model deviation indicators; Based on the model deviation index, the model parameters of each expert layer in step S1 are updated through the incremental learning algorithm, and the similarity threshold of the routing module in S3 and the weight distribution strategy of the decision fusion module in S5 are adjusted synchronously.

[0030] Model deviation index is generated by comparing the prediction results with the actual power generation measured value operation data, quantifying the model deviation and locating the light prediction floor height error module; Incremental learning updates parameters by fine-tuning only the meteorological processing layer with significant deviations, rather than global retraining, thus reducing computational overhead. The routing and fusion strategies are optimized in a coordinated manner, and the similarity threshold and fusion weight distribution of the routing module are adjusted synchronously to form a "lightweight iterative optimization" mode. While ensuring the dynamic evolution of the model, it avoids the waste of resources caused by large-scale repeated training and adapts to the high-frequency scheduling needs of virtual power plants.

[0031] In step S2, preprocessing includes: Normalize the temperature, humidity and wind speed in meteorological data; Perform time series alignment and missing value interpolation on electricity price data; The load data and light data are subjected to outlier removal and smoothing filtering respectively.

[0032] Meteorological data normalization is used to eliminate the scale differences of temperature and humidity parameters and avoid overfitting of the model to high-dimensional features; Electricity price time series alignment unifies the timestamps of discrete electricity price data and continuous load data, solving the problem of inconsistent sampling frequencies of multi-source data; Outlier removal and smoothing filtering are aimed at instantaneous noise in light sensors or occasional failures in load collection equipment, filtering out abnormal data points, improving the quality of input data, and reducing the interference of "dirty data" on model training through refined data cleaning, providing highly consistent input for subsequent stratified processing, and reducing fluctuations in prediction results.

[0033] In step S4, the local prediction results of the expert layer are further used to generate an intermediate decision report, which includes: Feature extraction paths and processing logic for each expert layer; Dynamic allocation records of routing modules; The contribution ratio of local prediction results to the final prediction is used to verify the model transparency.

[0034] Visualize the feature extraction path, record the feature processing logic of each expert layer, and support internal logic audit of the model; Route allocation record tracking, showing data distribution paths, and assisting in troubleshooting route anomalies; Contribution ratio analysis quantifies the contribution of each expert layer to the final prediction, provides a priority basis for model optimization, enhances dispatchers' trust in the prediction results by making the intermediate process transparent, and provides data support for model iteration.

[0035] The final energy forecast results include at least one of the following: Load demand curve for the next 24 hours; Probability distribution of photovoltaic power generation; Carbon emission sensitivity matrix, used to evaluate the carbon economy of different dispatch strategies.

[0036] The load demand curve provides hourly load forecasts for the next 24 hours, guiding virtual power plants to adjust energy storage charging and discharging strategies in real time; The probability distribution of power generation is within the confidence interval of the output photovoltaic power generation, with a 90% probability of power generation between 100-150MW, supporting risk-aware scheduling; The quantification of the carbon emission sensitivity matrix and the impact of different scheduling strategies on carbon emissions provide a quantitative basis for carbon economic optimization, convert the prediction results into scheduling instructions that can be directly implemented, open up the "prediction-decision-execution" link, and improve the operational efficiency of virtual power plants.

[0037] The training of the expert layer adopts a staged joint optimization strategy: In the first stage, each expert layer is trained independently and the model parameters are optimized using the historical data of the corresponding modality; In the second stage, the expert layer parameters are fixed, and the routing module and decision fusion module are trained end-to-end to minimize the overall prediction error; The third stage jointly fine-tunes all modules to improve the generalization ability of the model.

[0038] The first phase of independent training is to isolate the training process of each expert layer (for example, training LSTM with only meteorological data) to avoid interference between modal features and ensure basic performance. The second phase fixed-parameter training routing and fusion modules: Under the premise of stable expert layer, the routing distribution logic and fusion weight allocation strategy are optimized to improve the module collaboration efficiency; The third phase of joint fine-tuning involves fine-tuning all module parameters in an end-to-end manner, addressing local optimality and enhancing model generalization. This phased approach balances training efficiency and model accuracy, avoiding convergence difficulties caused by parameter confusion in the early stages of joint training while also reducing the risk of overfitting.

[0039] The method is deployed in the edge-cloud collaborative architecture: Edge devices perform data preprocessing, routing assignment, and expert-level local computation; The cloud server receives the local prediction results of each edge device and performs decision fusion and global optimization; The cloud sends the optimized scheduling instructions to the edge device to form a closed-loop control link.

[0040] Lightweight computing at the edge: Preprocessing, routing distribution, and expert-level prediction are performed on edge devices close to the data source (such as photovoltaic power plant controllers) to reduce network transmission latency. Cloud-based global optimization: Centrally handle high-density computing tasks (such as fusion decision-making and incremental learning), and leverage cloud computing power to generate complex strategies. Closed-Loop Command Delivery: Cloud-optimized dispatch instructions (such as energy storage charging and discharging plans) are delivered to edge devices in real time, forming a closed loop of "edge perception - cloud decision-making - edge execution." Benefit: By offloading computing tasks in a hierarchical manner, this system balances real-time performance with computational accuracy, adapting to the distributed and heterogeneous deployment characteristics of virtual power plants and reducing reliance on centralized servers.

[0041] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A full-modal hybrid expert model energy prediction method based on hierarchical routing, characterized by: The following steps are involved: S1. Constructing a hierarchical expert network structure, wherein the hierarchical expert network structure includes multiple expert layers, each expert layer is configured to process feature data of a specific modality, wherein the expert layers are divided into a meteorological processing layer, an electricity price processing layer, a load processing layer, and an illumination processing layer based on the modality type; S2. Receive multimodal feature data, including meteorological data, electricity price data, load data, and light data, and preprocess the data to generate a standardized time series data stream; S3. Input the standardized time series data stream into a routing module. The routing module dynamically allocates each modal data to a corresponding expert layer based on a feature similarity measure between a preset modal label and the input data. The output of the routing module is a mapping relationship between each modal data and its target expert layer. S4. In each expert layer, feature extraction and local prediction are performed on the assigned modal data to generate expert layer local prediction results, wherein the local prediction results include load prediction values, power generation prediction values, and carbon emission related indicators; S5. Receive the local prediction results of all expert layers through the decision fusion module, dynamically adjust the weights based on the historical prediction accuracy of each expert layer, use the attention mechanism to perform weighted fusion on the local prediction results, and output the final energy prediction result; S6. Feedback the energy forecast results to the virtual power plant dispatching system for real-time adjustment of power generation plans, load distribution or carbon emission optimization strategies.

2. The energy forecasting method based on hierarchical routing and full modal hybrid expert model according to claim 1 is characterized by: In step S1, the expert layer adopts a heterogeneous model structure, including at least one of the following combinations: the meteorological processing layer adopts a long short-term memory neural network, the electricity price processing layer adopts a gradient boosting decision tree, the load processing layer adopts a convolutional neural network, and the illumination processing layer adopts support vector regression.

3. The energy forecasting method based on hierarchical routing and full modal hybrid expert model according to claim 1 is characterized by: In step S3, the dynamic allocation logic of the routing module includes: Calculate the cosine similarity between the feature vector of the input data and the preset modal labels of each expert layer; Select the expert layer with the highest similarity as the target routing path; If the similarity is lower than the preset threshold, the exception handling mechanism is triggered and the data is routed to the default expert layer for cross-modal joint analysis.

4. The energy forecasting method based on hierarchical routing and full modal hybrid expert model according to claim 1 is characterized by: In step S5, the attention mechanism includes the following steps: Generate initial weights based on the historical prediction errors of each expert layer; Dynamically scoring the confidence of the local prediction results through a multi-layer perceptron to correct the initial weights; The local prediction results are weighted and summed according to the corrected weights to generate the final prediction result.

5. The energy prediction method based on hierarchical routing and full modal hybrid expert model according to claim 1 is characterized by: In step S5, the decision fusion module includes the following steps: Based on the local prediction results of each expert layer generated by S4, the historical prediction error records of each expert layer are extracted to generate the initial weight distribution table; Associating the data mapping relationship output by the routing module in S3 with the initial weight distribution table, dynamically scoring the confidence of the local prediction results through a multi-layer perceptron, and correcting the initial weights; The local prediction results of S4 are weighted and fused according to the revised weights to generate the final energy prediction results; Receive real-time operational data from S6's virtual power plant dispatching system, including actual load values, power generation values, and carbon emission monitoring results; Performing an error comparison between the operating data and the final energy forecast result to generate a model deviation index; Based on the model deviation index, the model parameters of each expert layer in step S1 are updated through the incremental learning algorithm, and the similarity threshold of the routing module in S3 and the weight allocation strategy of the decision fusion module in S5 are adjusted synchronously.

6. The energy prediction method based on hierarchical routing and full modal hybrid expert model according to claim 1 is characterized by: In step S2, the pre-processing includes: Normalize the temperature, humidity and wind speed in meteorological data; Perform time series alignment and missing value interpolation on electricity price data; The load data and light data are subjected to outlier removal and smoothing filtering respectively.

7. The energy forecasting method based on hierarchical routing and full modal hybrid expert model according to claim 1 is characterized by: In step S4, the local prediction results of the expert layer are further used to generate an intermediate decision report, which includes: Feature extraction paths and processing logic for each expert layer; Dynamic allocation records of routing modules; The contribution ratio of local prediction results to the final prediction is used to verify the model transparency.

8. The energy forecasting method based on hierarchical routing and full modal hybrid expert model according to claim 1 is characterized by: The final energy forecast result includes at least one of the following: Load demand curve for the next 24 hours; Probability distribution of photovoltaic power generation; Carbon emission sensitivity matrix, used to evaluate the carbon economy of different dispatch strategies.

9. The energy forecasting method based on hierarchical routing and full modal hybrid expert model according to claim 1 is characterized by: The training of the expert layer adopts a staged joint optimization strategy: In the first stage, each expert layer is trained independently and the model parameters are optimized using the historical data of the corresponding modality; In the second stage, the expert layer parameters are fixed, and the routing module and decision fusion module are trained end-to-end to minimize the overall prediction error; The third stage jointly fine-tunes all modules to improve the generalization ability of the model.

10. The energy prediction method based on hierarchical routing and full modal hybrid expert model according to claim 1 is characterized by: The method is deployed in the edge-cloud collaborative architecture: Edge devices perform data preprocessing, routing assignment, and expert-level local computation; The cloud server receives the local prediction results of each edge device and performs decision fusion and global optimization; The cloud sends the optimized scheduling instructions to the edge device to form a closed-loop control link.

Images