Shared timing diagram network and multi-task learning method and system for household light storage

Abstract

The present application relates to the technical field of household photovoltaic energy storage collaborative control, in particular to a household photovoltaic storage collaborative method and system based on shared timing diagram network and multi-task learning, which first collects multi-source heterogeneous data to construct a household micro-grid node space-time topology graph, uses a space-time enhancement transformer as a unified shared backbone network to extract global shared space-time features; based on the shared features, multi-task joint decoding is performed to output photovoltaic / load prediction values, SOC trajectories and confidence intervals, and same variance uncertainty is introduced to dynamically adjust the task loss weights; the prediction results are input into a model predictive controller with fused physical hard constraints to solve optimal charging / discharging instructions; finally, through knowledge distillation by adding a physical consistency penalty term, a cloud large model is compressed into an edge lightweight model to realize local safe autonomous scheduling; the present application solves problems such as prediction error cascade amplification, weak resistance to extreme electricity prices, and easy triggering of battery physical over-limit by AI black box, and significantly improves the operation economy and control safety of the system.

Description

Technical Field

[0001] This invention relates to the field of collaborative control technology for residential photovoltaic energy storage, specifically to a collaborative method and system for residential photovoltaic energy storage using shared time-series graph networks and multi-task learning. Background Technology

[0002] With the increasing prevalence of residential photovoltaic (PV) and energy storage, home microgrids are evolving from passive response to active participation in grid interaction. Especially in market environments implementing time-of-use (TOU) pricing or even real-time spot pricing, accurately predicting future PV output and household load, and combining this with dynamic pricing to formulate optimal charging and discharging strategies, has become a core competitive advantage for home energy storage systems. Currently, the industry mainly adopts a "step-by-step forecasting and heuristic rule / conventional MPC" approach. For example, using LSTM or LightGBM models to establish independent PV forecasting and load forecasting models, and then inputting the point forecasting results into the rule-based model.

[0003] Existing solutions have many shortcomings. The spatiotemporal coupling of multi-source heterogeneous data is missing. Traditional models independently predict photovoltaics and loads, ignoring the spatial topological relationships between household devices and external weather. Each prediction module is independent and does not extract shared implicit features, resulting in prediction errors being amplified by superposition at the scheduling layer.

[0004] Existing strategies are mostly based on deterministic forecasts and static electricity price rules. When faced with dynamic spot electricity prices, they are prone to economic losses or frequent charging and discharging malfunctions. Furthermore, traditional models only output a definite forecast value and do not provide a confidence interval for the forecast. When high volatility in electricity prices occurs, the dispatcher cannot assess the risks of aggressive arbitrage.

[0005] Meanwhile, there is a contradiction between large cloud models and low edge computing power. Large models have a large number of parameters and are difficult to deploy directly on residential energy storage edge gateways with limited computing power. Purely data-driven AI models lack physical perception and are prone to violating physical laws when performing SOC estimation. Furthermore, the fixed-weight loss function in multi-task training often leads to overfitting of easy tasks and failure to converge for difficult tasks, resulting in gradient conflicts that make it difficult to achieve optimal convergence. Therefore, in view of the above situation, there is an urgent need to develop a collaborative method and system for residential photovoltaic energy storage with shared temporal graph networks and multi-task learning to overcome the shortcomings in current practical applications. Summary of the Invention

[0006] The purpose of this invention is to provide a user-friendly optical-storage collaborative method and system for shared temporal graph networks and multi-task learning, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A collaborative approach for residential optical-storage systems using shared temporal graph networks and multi-task learning includes the following steps: S1. Collect multi-source heterogeneous data and construct a spatiotemporal topology map of home microgrid nodes; S2. Input the feature map sequence within the continuous time window into the spatiotemporal enhancement transformer to extract the globally shared spatiotemporal feature tensor; S3. Perform multi-task joint decoding based on the shared spatiotemporal feature tensor, and output photovoltaic prediction value, load prediction value, SOC estimated trajectory and corresponding confidence interval in parallel. S4. Input the predicted value, SOC estimated trajectory and confidence interval into the model prediction controller, and combine the battery physical hard constraints and dynamic electricity price to solve the optimal charge and discharge command in a rolling manner. S5. By adding a physical consistency penalty term to knowledge distillation, the large cloud model is compressed into a lightweight edge model and local secure autonomous scheduling is performed.

[0008] As a further aspect of the present invention: in step S2, the feature extraction process of the spatiotemporal enhancement transformer includes: The spatial attention-weighted feature representation is calculated based on the GAT graph attention model, using the following formula: in, This is the feature representation after spatial attention weighting; These are the learnable weight projection matrices for the query, key, and value, respectively. This is a scaling factor for the feature dimension, used to prevent the dot product from becoming too large; For prior adjacency matrix The constructed topological mask matrix; After adding the spatial attention-weighted feature representation to the temporal location encoding, it is input into a multi-layer Transformer Encoder to extract the globally shared spatiotemporal feature tensor.

[0009] As a further aspect of the present invention: in step S3, the multi-task joint decoding includes: A multilayer sensor is used as the prediction head to map the output photovoltaic predicted power and load predicted power. The basic reference value is calculated using an equivalent circuit model and ampere-hour integration. The nonlinear residual is corrected by the gated cyclic unit, and the SOC estimate is output. Quantile regression is used as the uncertainty head, and the upper and lower bounds of the prediction are output at a set confidence level.

[0010] As a further aspect of the present invention: In step S3, during the model training phase, homoscedastic uncertainty is used to dynamically adjust the loss weights of each task, and the total loss function for joint training is: ; in, This represents the total number of tasks. For the first The original loss of each task; For the first Learnable noise parameters for each task.

[0011] As a further aspect of the present invention: in step S4, the rolling optimization objective function constructed by the model prediction controller is: ; in, To predict the number of steps; For the future Electricity price at any given time; The power interacting with the power grid; For time step; Weighting for penalties related to battery life loss; Battery charging and discharging power The resulting equivalent cyclic loss function; The optimization process satisfies the following hard physical constraints: in, For the future state of charge; These refer to the battery charging and discharging efficiency, respectively. For charging and discharging control power commands; This refers to the battery's rated total capacity.

[0012] Simultaneously, boundary constraints are considered: and .

[0013] As a further aspect of the present invention: in step S5, the total loss function of knowledge distillation is: in, For the student model, use hard loss based on the true labels; Distillation of knowledge loss to approximate the soft labels of the teacher model in the student model; The SOC value predicted by the student model; The theoretical SOC value is calculated based on the physical law of ampere-hour integration. The set hyperparameter weights.

[0014] A shared time-series graph network and multi-task learning-based residential photovoltaic-storage collaborative control system includes: The data acquisition and topology construction module is used to collect multi-source heterogeneous data and construct a spatiotemporal topology map of home microgrid nodes. The spatiotemporal feature encoding module has a built-in spatiotemporal enhancement transformer, which is used to input the feature map sequence within a continuous time window into the spatiotemporal enhancement transformer to extract the globally shared spatiotemporal feature tensor; The multi-task joint decoding and training module includes multiple parallel task output heads, which are used to input the shared spatiotemporal feature tensor into the multiple parallel task output heads and output photovoltaic prediction values, load prediction values, SOC estimation trajectories and corresponding confidence intervals in parallel. The physical constraint optimization scheduling module has a built-in model prediction controller, which is used to input the predicted value, SOC estimated trajectory and confidence interval into the model prediction controller, and combine the battery physical hard constraints and dynamic electricity price to solve the optimal charge and discharge command in a rolling manner. The cloud-edge collaborative distillation deployment module is used to compress large cloud models into lightweight edge models by adding physical consistency penalty terms for knowledge distillation, and to perform local secure autonomous scheduling.

[0015] As a further aspect of the present invention: the spatiotemporal enhancement transformer includes a spatial feature extraction unit based on the GAT graph attention model and a temporal feature extraction unit based on a multi-layer Transformer Encoder.

[0016] As a further aspect of the present invention: the multi-task joint decoding and training module has a built-in multi-task adaptive weight training mechanism, which dynamically adjusts the loss weights of each task using homoscedastic uncertainty.

[0017] As a further aspect of the present invention, the cloud-edge collaborative distillation deployment module also includes a network outage autonomous unit, which is used to control the edge gateway to achieve local secure autonomous scheduling using the lightweight model and physical consistency module when cloud communication is interrupted.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the problem of missing spatiotemporal coupling in multi-source heterogeneous data. By constructing a spatiotemporal topology map of home microgrid nodes, it establishes spatial connections between devices and the external environment. It uses a spatiotemporal enhancement converter as a unified shared backbone network to extract the temporal evolution patterns and spatial coupling characteristics of multidimensional variables, thus avoiding the cascading amplification of errors caused by independent calculations of multiple models and reducing the waste of computing power. This enhances the system's resilience under extreme electricity prices. Based on shared features, it performs multi-task joint decoding and outputs the predicted values ​​and confidence intervals of photovoltaic power output, household load, and battery SOC in parallel, providing a risk assessment basis for scheduling decisions. At the same time, it introduces homoscedastic uncertainty to dynamically adjust the loss weights of each task, solving the gradient conflict problem in multi-task training and achieving optimal convergence of the model. It effectively avoids the problem of battery physical limits. By integrating model predictive control with physical hard constraints to optimize scheduling, it transforms AI prediction output into charging and discharging commands that conform to the laws of electrical physics and have the best overall economic benefits. It achieves dynamic electricity price arbitrage benefits while taking into account the long service life of the battery. This solves the problem of limited computing power when deploying large models at the edge. By adding a physical consistency penalty term to the knowledge distillation method, the large cloud model is compressed into a lightweight edge model, which forces the edge model output to conform to physical laws and ensures local secure autonomous scheduling at the edge when cloud communication is interrupted. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the user optical-storage collaborative method using a shared temporal graph network and multi-task learning in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0022] Please see Figure 1 The shared time-series graph network and multi-task learning method and system for collaborative home optical-storage systems provided in this invention construct a unified home optical-storage collaborative control system architecture consisting of a shared time-series backbone, a multi-task output head, a security constraint scheduler, and cloud-edge collaborative deployment.

[0023] First, a spatiotemporal topology map of the home microgrid is constructed by collecting heterogeneous data from multiple sources, establishing spatial connections between devices and the external environment. Then, a spatiotemporal enhancement converter is used as a unified shared backbone network, while extracting the temporal evolution patterns and spatial coupling characteristics of multidimensional variables to avoid the cascading amplification of errors caused by independent calculations of multiple models. Next, multi-task joint decoding is performed based on shared features, outputting the predicted values ​​and confidence intervals of photovoltaic power output, household load, and battery SOC in parallel, and introducing homoscedastic uncertainty to dynamically adjust the loss weights of each task to solve the gradient conflict problem in multi-task training. Subsequently, the prediction results are integrated with the dynamic electricity price input into a model prediction controller with physical hard constraints, and the optimal charging and discharging command with comprehensive economic benefits is solved in a rolling manner. Finally, by adding a knowledge distillation method with a physical consistency penalty term, the large cloud model is compressed into a lightweight edge model to achieve local secure autonomous scheduling in the event of a network outage.

[0024] This invention can effectively solve the problems of prediction lag, data silos among multiple devices, and battery physical over-limits caused by AI black box control in existing residential photovoltaic energy storage systems under highly volatile external environments, and significantly improve the system's risk resistance and operational economy.

[0025] Example 1: A collaborative control method for residential photovoltaic and energy storage systems based on shared temporal graph networks and multi-task learning This embodiment provides a collaborative control method for residential photovoltaic-storage systems based on shared time-series graph networks and multi-task learning. The core of this method lies in constructing a unified technical architecture consisting of a shared time-series backbone, multi-task output heads, a safety constraint scheduler, and cloud-edge collaborative deployment. This architecture effectively addresses issues in existing residential photovoltaic-storage systems under highly volatile external environments, such as prediction lag, data silos among multiple devices, and the potential for battery physical limit violations caused by AI black-box control. The method specifically includes the following steps: Step S1: Multi-source heterogeneous data acquisition and spatiotemporal topology graph construction The raw data sequence is collected in 15-minute time steps Δt. The raw data sequence includes photovoltaic inverter power, total household load of smart meters, battery voltage / current / temperature collected by BMS, irradiance / cloud cover forecast data issued by the meteorological bureau, and spot / time-of-use electricity price data.

[0026] Based on the collected multi-source heterogeneous data, a node feature matrix is ​​constructed. and the prior adjacency matrix reflecting the electrical relationships between equipment Where N is the total number of nodes in the defined graph network, including photovoltaic nodes, load nodes, and energy storage nodes; D is the feature dimension contained in each node at a single time step; Let be the feature set of all nodes at time t; A is a binary or continuous weight matrix initialized based on physical connections and correlations.

[0027] This step establishes spatial topological relationships between physical devices within a home microgrid and between devices and the external environment by unifying multi-source data formats. This provides high-quality and structured initial features for subsequent models and solves the problem that traditional models ignore spatial topological relationships between devices and weather conditions.

[0028] Step S2: Unified Feature Encoding Based on Spatiotemporal Enhancement Transformer (ST-Transformer) The feature map sequence within a continuous time history window T The topological adjacency matrix A is input to the spatiotemporal enhancement transformer to extract the shared hidden layer temporal feature tensor that incorporates spatiotemporal dependencies. .

[0029] The specific implementation process includes: 1. Spatial Feature Extraction: Calculate spatial attention-weighted feature representations based on the GAT graph attention model. The calculation formula is: in, This is the feature representation after spatial attention weighting; These are the learnable weight projection matrices for the query, key, and value, respectively. This is a scaling factor for the feature dimension, used to prevent the dot product from becoming too large; For prior adjacency matrix The constructed topology mask matrix has a mask value of - for nodes that are not physically related. .

[0030] 2. Temporal Feature Extraction: ... After incorporating temporal location encoding, the data is input into a multi-layer TransformerEncoder to extract global features. .

[0031] This step uses a single shared backbone network to simultaneously extract the evolution patterns of multidimensional variables over time and their coupling relationships in spatial topology, reducing the waste of computing power caused by independent calculations of multiple models, while avoiding the defect of error cascading amplification in traditional step-by-step prediction methods.

[0032] Step S3: Dynamic electricity price perception and multi-task joint decoding and dynamic weight training Shared feature tensor The system takes multiple parallel task output heads as input, performs multi-task joint decoding, and outputs the photovoltaic / load forecast values, SOC trajectory, and confidence intervals for the next H time periods. Simultaneously, homoscedastic uncertainty is introduced during the model training phase to dynamically adjust the loss weights of each task, avoiding gradient dominance.

[0033] The specific implementation process includes: 1. Photovoltaic and Load Forecasting: A multilayer sensor (MLP) is used as the forecast head to map and output the predicted power. and ; 2. SOC estimation and trajectory prediction: The basic reference value is calculated using the equivalent circuit model (ECM) and ampere-hour integration. The nonlinear residual is corrected by the gated cyclic unit (GRU) and the output SOC(t) is obtained. 3. Dynamic Electricity Price Perception and Uncertainty Quantification: QRNN quantile regression is used as the uncertainty head, outputting the upper and lower bounds of the prediction at confidence level α. Its quantile loss function is: ; in, This is the quantile regression loss; For the set target quantile (e.g.) It means 90% ), For real labels Compared with model predictions The residuals between; 4. Multi-task adaptive weight training: The total loss function during joint training. Designed as follows: ; in, The total number of tasks (e.g., photovoltaic, load, SOC, etc.); For the first The original loss of each task (such as MSE or Huber Loss); For the first Learnable noise parameters (homoscedastic uncertainty) for each task.

[0034] When the error of a certain task is extremely large, the model adaptively increases. This reduces the weight of the task in the current iteration step. This is to prevent it from disrupting parameter updates on the shared backbone. As a regularization term to prevent It can grow indefinitely.

[0035] This step achieves parallel prediction of photovoltaic, load and SOC through multi-task joint decoding, and introduces uncertainty quantification to provide risk assessment basis for scheduling decisions; it adopts homoscedastic uncertainty to dynamically adjust the loss weight, solves the gradient conflict problem in multi-task training, and achieves optimal convergence of the model.

[0036] Step S4: Model Predictive Control (MPC) Optimization Scheduling Incorporating Physical Constraints The forecast, SOC trajectory, confidence interval, and dynamic electricity price will be used. By inputting the equipment constraint parameters into the model predictive controller, a rolling optimization objective function is constructed within the finite prediction time domain H. The optimal control command, including charging power, is then obtained by solving the problem. With discharge power .

[0037] The objective function for the rolling optimization is: ; in, For predicting steps (such as look-ahead) Hours, Step Size minutes, then ); For the future Electricity price at any given time; This represents the power exchanged with the power grid (positive for purchasing electricity, negative for selling electricity). For time step; Weighting for penalties related to battery life loss; Battery charging and discharging power The resulting equivalent cyclic loss function.

[0038] The optimization process must satisfy the following underlying physical constraints: in, For the future state of charge; These refer to the battery charging and discharging efficiency, respectively. For charging and discharging control power commands; This refers to the battery's rated total capacity.

[0039] Simultaneously, boundary constraints are considered: and (Cannot be charged and discharged simultaneously).

[0040] This step transforms the AI ​​prediction output into execution instructions that conform to the laws of electrophysics and have the best overall economic benefits. By introducing a battery life loss penalty, it achieves electricity price arbitrage while taking into account the long-term lifespan of the battery. Through hard physical constraints, it effectively avoids the battery physical limit exceeding problem that is easily caused by pure data-driven AI models.

[0041] Step S5: Cloud-edge collaborative distillation deployment based on physical consistency The large model trained in the cloud is used as the teacher model. Based on the computing power constraints of edge devices, a local lightweight student model is obtained by compressing it through knowledge distillation. This model is then deployed on a residential energy storage edge gateway (usually an ARM architecture industrial control computer) to perform local control.

[0042] The distillation loss function not only includes the traditional output matching loss, but also rigidly incorporates a physical consistency penalty term, specifically: in, For the student model, use hard loss based on the true labels; Distillation of knowledge loss to approximate the soft labels of the teacher model in the student model; The SOC value predicted by the student model; The theoretical SOC value is calculated based on the physical law of ampere-hour integration. The set hyperparameter weights.

[0043] This step addresses the limitation of computing power in deploying large models at the edge. Simultaneously, by using a physical consistency penalty term, it forces the output of the lightweight edge AI model to conform to physical laws, ensuring the control security of the edge during network outages. When cloud communication is interrupted, the edge gateway can utilize the degraded lightweight model and the physical consistency module to achieve local secure autonomous scheduling.

[0044] Example 2: A Residential Photovoltaic-Storage Collaborative Control System Based on Shared Time Sequence Graph Networks and Multi-Task Learning This embodiment provides a user-friendly optical-storage collaborative control system based on shared time-series graph networks and multi-task learning, used to implement the method described in Embodiment 1 above. The system includes: The data acquisition and topology construction module is used to acquire raw data sequences at 15-minute time steps Δt. These raw data sequences include photovoltaic inverter power, total household load from smart meters, battery voltage / current / temperature collected by the BMS, irradiance / cloud cover forecast data from the meteorological bureau, and spot / time-of-use electricity price data. Based on the acquired multi-source heterogeneous data, a node feature matrix is ​​constructed. and the prior adjacency matrix reflecting the electrical relationships between equipment .

[0045] Spatiotemporal feature encoding module: Built-in spatiotemporal enhancement transform (ST-Transformer) is used to transform the feature map sequence within a continuous time history window T. The topological adjacency matrix A is input to the spatiotemporal enhancement transformer to extract the shared hidden layer temporal feature tensor that incorporates spatiotemporal dependencies. .

[0046] The spatiotemporal enhancement transformer includes a spatial feature extraction unit based on the GAT graph attention model and a temporal feature extraction unit based on a multi-layer Transformer Encoder.

[0047] Multi-task joint decoding and training module: includes multiple parallel task output heads, namely photovoltaic and load forecasting head, SOC estimation and trajectory prediction head, dynamic electricity price sensing and uncertainty head; used to share feature tensors The module takes multiple parallel task output headers as input and outputs the photovoltaic / load forecasts, SOC trajectories, and confidence intervals for the next H time periods. Simultaneously, it incorporates a multi-task adaptive weight training mechanism, dynamically adjusting the loss weights of each task using homoscedastic uncertainty.

[0048] Physical constraint optimization scheduling module: Includes a built-in Model Prediction Controller (MPC) to integrate predicted values, SOC trajectory, confidence intervals, and dynamic electricity prices. By inputting the equipment constraint parameters into the model predictive controller, a rolling optimization objective function is constructed within the finite prediction time domain H. The optimal control command, including charging power, is then obtained by solving the problem. With discharge power This module incorporates a battery physical hard constraint check unit to ensure that the output control commands comply with the laws of electrophysics.

[0049] The cloud-edge collaborative distillation deployment module uses a large model trained in the cloud as the teacher model. Based on the computing power constraints of edge devices, it compresses the large model into a local lightweight student model using a knowledge distillation method that incorporates a physical consistency penalty term. This student model is then deployed on the residential energy storage edge gateway to perform local control. This module also includes a network outage autonomous unit, which controls the edge gateway to achieve local secure autonomous scheduling using the degraded lightweight model and the physical consistency module when cloud communication is interrupted.

[0050] Optionally, the system also includes a data storage module for storing collected historical data, model training parameters, and scheduling execution records, providing data support for continuous model optimization.

[0051] It should be noted that, in this invention, although the specification describes the embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A collaborative method for user-based optical-storage systems using shared temporal graph networks and multi-task learning, characterized in that: Includes the following steps: S1. Collect multi-source heterogeneous data and construct a spatiotemporal topology map of home microgrid nodes; S2. Input the feature map sequence within the continuous time window into the spatiotemporal enhancement transformer to extract the globally shared spatiotemporal feature tensor; S3. Perform multi-task joint decoding based on the shared spatiotemporal feature tensor, and output photovoltaic prediction value, load prediction value, SOC estimated trajectory and corresponding confidence interval in parallel. S4. Input the predicted value, SOC estimated trajectory and confidence interval into the model prediction controller, and combine the battery physical hard constraints and dynamic electricity price to solve the optimal charge and discharge command in a rolling manner. S5. By adding a physical consistency penalty term to knowledge distillation, the large cloud model is compressed into a lightweight edge model and local secure autonomous scheduling is performed.

2. The method for collaborative operation of user optical storage with shared temporal graph networks and multi-task learning according to claim 1, characterized in that, In step S2, the feature extraction process of the spatiotemporal enhancement transformer includes: The spatial attention-weighted feature representation is calculated based on the GAT graph attention model, using the following formula: in, This is the feature representation after spatial attention weighting; These are the learnable weight projection matrices for the query, key, and value, respectively. This is a scaling factor for the feature dimension, used to prevent the dot product from becoming too large; For prior adjacency matrix The constructed topological mask matrix; After adding the spatial attention-weighted feature representation to the temporal location encoding, it is input into a multi-layer TransformerEncoder to extract the globally shared spatiotemporal feature tensor.

3. The method for collaborative operation of user optical storage with shared temporal graph networks and multi-task learning according to claim 1, characterized in that, In step S3, the multi-task joint decoding includes: A multilayer sensor is used as the prediction head to map the output photovoltaic predicted power and load predicted power. The basic reference value is calculated using an equivalent circuit model and ampere-hour integration. The nonlinear residual is corrected by the gated cyclic unit, and the SOC estimate is output. Quantile regression is used as the uncertainty head, and the upper and lower bounds of the prediction are output at a set confidence level.

4. The method for collaborative operation of user optical storage with shared temporal graph networks and multi-task learning according to claim 1, characterized in that, In step S3, during the model training phase, homoscedastic uncertainty is used to dynamically adjust the loss weights of each task. The total loss function for joint training is: in, This represents the total number of tasks. For the first The original loss of each task; For the first Learnable noise parameters for each task.

5. The method for collaborative operation of user optical storage with shared temporal graph networks and multi-task learning according to claim 1, characterized in that, In step S4, the rolling optimization objective function constructed by the model prediction controller is: in, To predict the number of steps; For the future Electricity price at any given time; The power interacting with the power grid; For time step; Weighting for penalties related to battery life loss; Battery charging and discharging power The resulting equivalent cyclic loss function; The optimization process satisfies the following hard physical constraints: in, For the future state of charge; These refer to the battery charging and discharging efficiency, respectively. For charging and discharging control power commands; This refers to the battery's rated total capacity. Simultaneously, boundary constraints are considered: and .

6. The method for collaborative operation of user optical storage with shared temporal graph networks and multi-task learning according to claim 1, characterized in that, In step S5, the total loss function of knowledge distillation is: in, For the student model, use hard loss based on the true labels; Distillation of knowledge loss to approximate the soft labels of the teacher model in the student model; The SOC value predicted by the student model; The theoretical SOC value is calculated based on the physical law of ampere-hour integration. The set hyperparameter weights.

7. A residential optical-storage collaborative control system based on shared time-series graph networks and multi-task learning, characterized in that, include: The data acquisition and topology construction module is used to collect multi-source heterogeneous data and construct a spatiotemporal topology map of home microgrid nodes. The spatiotemporal feature encoding module has a built-in spatiotemporal enhancement transformer, which is used to input the feature map sequence within a continuous time window into the spatiotemporal enhancement transformer to extract the globally shared spatiotemporal feature tensor; The multi-task joint decoding and training module includes multiple parallel task output heads, which are used to input the shared spatiotemporal feature tensor into the multiple parallel task output heads and output photovoltaic prediction values, load prediction values, SOC estimation trajectories and corresponding confidence intervals in parallel. The physical constraint optimization scheduling module has a built-in model prediction controller, which is used to input the predicted value, SOC estimated trajectory and confidence interval into the model prediction controller, and combine the battery physical hard constraints and dynamic electricity price to solve the optimal charge and discharge command in a rolling manner. The cloud-edge collaborative distillation deployment module is used to compress large cloud models into lightweight edge models by adding physical consistency penalty terms for knowledge distillation, and to perform local secure autonomous scheduling.

8. The shared time-series graph network and multi-task learning-based residential optical-storage collaborative control system according to claim 7, characterized in that, The spatiotemporal enhancement transformer includes a spatial feature extraction unit based on the GAT graph attention model and a temporal feature extraction unit based on a multi-layer TransformerEncoder.

9. The shared time-series graph network and multi-task learning-based residential optical-storage collaborative control system according to claim 7, characterized in that, The multi-task joint decoding and training module has a built-in multi-task adaptive weight training mechanism, which dynamically adjusts the loss weights of each task using homoscedastic uncertainty.

10. The shared time-series graph network and multi-task learning-based residential optical-storage collaborative control system according to claim 7, characterized in that, The cloud-edge collaborative distillation deployment module also includes a network outage autonomous unit, which is used to control the edge gateway to achieve local secure autonomous scheduling using the lightweight model and physical consistency module when cloud communication is interrupted.

Images