Optical storage system adaptive scheduling method based on multi-modal data fusion

Abstract

The invention relates to the technical field of optical storage scheduling, and discloses an optical storage system adaptive scheduling method based on multi-modal data fusion, and the method comprises the following steps: S1, collecting structured static data, time sequence monitoring dynamic data, external market data, and image data; s2, obtaining a space-time coupling feature matrix, a solar radiation predicted value and a power grid load predicted value, and constructing a multi-modal feature matrix through the space-time coupling feature matrix, the solar radiation predicted value, the power grid load predicted value, the structured static data and the external market data; s3, generating a random control strategy through the multi-modal data feature matrix and the three independent agents, optimizing the random control strategy to obtain a deterministic control strategy, and outputting the deterministic control strategy through a reinforcement learning controller; and S4, repeating the steps S1 to S3, and stably outputting a corresponding deterministic control strategy at each time step by the optical storage system. According to the invention, efficient response to complex environments and uncertain events can be realized, and the scheduling precision of the optical storage system is improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic-storage scheduling technology, and specifically to an adaptive scheduling method for photovoltaic-storage systems based on multimodal data fusion. Background Technology

[0002] With the rapid development of renewable energy globally, photovoltaic (PV) power generation has become one of the important sources of clean energy. However, the volatility and uncertainty of PV power generation pose significant challenges to the stable operation of the power grid. Especially in environments with frequent changes in sunlight and large fluctuations in load demand, traditional PV-storage system scheduling methods often rely on fixed rules and conventional optimization algorithms, which cannot effectively cope with the volatility of PV power generation, changes in grid load, and the nonlinear characteristics of energy storage systems. This results in low scheduling accuracy, poor system stability, and a lack of online adaptive optimization capabilities, leading to increased load fluctuations in the power grid and making it difficult to achieve efficient energy consumption. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, the present invention aims to provide an adaptive scheduling method for photovoltaic storage systems based on multimodal data fusion, which can efficiently respond to complex environments and uncertain events, and improve the scheduling accuracy, stability and adaptability of photovoltaic storage systems.

[0004] The technical solution adopted in this invention is as follows: An adaptive scheduling method for a photoelectric storage system based on multimodal data fusion, comprising the following steps: S1: Collect structured static data, time-series monitoring dynamic data, and external market data through corresponding sensors deployed in photovoltaic power generation systems, energy storage devices, and power grid systems, as well as image data of the photovoltaic panel itself and its surrounding physical environment through visual sensors. S2: The spatiotemporal coupling feature matrix and solar radiation prediction value are obtained from the time-series monitoring dynamic data and image data at each time step, respectively. At the same time, a calendar feature matrix is ​​constructed. A short-term load forecasting model based on Bi-LSTM is established using the time-series monitoring dynamic data and calendar feature matrix at each time step to obtain the power grid load forecast value. Then, a multimodal feature matrix is ​​constructed using the spatiotemporal coupling feature matrix, solar radiation prediction value, power grid load forecast value, structured static data, and external market data at each time step. S3: Construct a standard state space for three independent agents using the multimodal data feature matrix. The three agents generate random actions through an actor network and allow the agents to select actions at corresponding time steps with a probability density that follows a Gaussian distribution. Generate a stochastic control policy. Optimize the stochastic control policy using a proximal policy optimization algorithm through a shared actor network and independent commentator networks to obtain a deterministic control policy. Output the deterministic control policy through a reinforcement learning controller. S4: Repeat S1-S3, and the optical storage system stably outputs the corresponding deterministic control strategy at each time step.

[0005] In a preferred embodiment of the present invention, in S1, structured static data is used to define the boundary conditions of the system, including the maximum charging and discharging power of the energy storage battery and the peak load limit allowed by the power grid. The time-series monitoring dynamic data includes light intensity, ambient temperature, relative humidity, real-time photovoltaic power generation, grid node load, and real-time state of charge of energy storage batteries; External market data includes real-time time-of-use electricity prices; A combination of a wide-angle lens and a fisheye lens is used to capture images of the photovoltaic panel itself and its surrounding physical environment.

[0006] In a preferred embodiment of the present invention, S2, obtaining the spatiotemporal coupling feature matrix and the solar radiation prediction value by monitoring time-series dynamic data and image data respectively, specifically includes the following steps: S211: Preprocess the acquired images and divide them into... Each local feature is treated as a node in a graph structure. Edges are constructed based on the adjacency relationships of the features in physical space. A multi-layer graph attention network is used to iteratively update the node features. After the layers are aggregated, a set of feature vectors containing global spatial occlusion information is obtained, as shown in formulas (1) and (2). (1), (2), (3), In formulas (1), (2) and (3), For the first After layer attention aggregation, the updated result is the first layer. The new feature vector of each node This is the layer index of the graph attention network. It is a non-linear activation function. For nodes neighborhood set, Attention coefficient For the first The learnable linear transformation weight matrix of the layer, For the first Layer nodes eigenvectors, It is the spatial characteristic matrix; S212: The spatial feature matrix is ​​linearly transformed to obtain the query, key, and value matrix. The query, key, and value matrix is ​​then input into the ST-Transformer model. The standard self-attention method is used to output the spatiotemporal feature matrix with spatiotemporal dependence, as shown in formulas (4), (5), (6), and (7). (4), (5), (6), (7), In formulas (4), (5), (6), (7), These are query, key, and value matrices, respectively. For spatial characteristic matrix, These are the learnable linear mapping matrices, Let T be the spatiotemporal feature matrix, and T be the time step. The dimension of the key vector; S213: The spatiotemporal feature matrix and the spatial feature matrix are fused along the channel dimension through a feature splicing layer to preserve local details and global temporal patterns, as shown in formula (8). (8), In formula (8), This is the characteristic matrix of environmental physical constraints. This is a global pooling operation. For spatial characteristic matrix, It is the spatiotemporal feature matrix; S214: Extract the feature matrix of the time-series monitoring dynamic data, and fuse the feature matrix of the time-series monitoring dynamic data and the feature matrix of environmental physical constraints through a gating unit to obtain dual modal information of visual and numerical values, as shown in formula (9). (9), In formula (9), The spatiotemporal coupling characteristic matrix, For adaptive gating weight coefficients, This is the characteristic matrix of environmental physical constraints. For time series feature extraction function, For time-series monitoring of dynamic data feature matrix; S215: Based on the spatiotemporal coupling feature matrix, the CNN-LSTM architecture is used to extract local spatiotemporal features, and a nonlinear autoregressive network is combined to capture the nonlinear relationship of the time series of solar radiation, as shown in formulas (10) and (11). (10) (11), In formula (10), For external input variables, For temperature, For humidity, For the current moment The previous Weather conditions at each time step, for index; In formula (11), For a moment The predicted value of solar radiation, and These are the mapping functions for nonlinear autoregressive networks and CNN-LSTM architectures, respectively. To delay the step size, These are measured values ​​of radiation at historical moments.

[0007] In a preferred embodiment of the present invention, S2, obtaining the power grid load forecast value through time-series monitoring dynamic data and calendar feature matrix specifically includes the following steps: S221: Construct the calendar feature matrix to provide temporal context information for the Bi-LSTM model, as shown in formula (12). (12) In formula (12), For the calendar feature matrix, This represents the hourly index of the current time within a day. Workday type identifier Indicates the sign for statutory holidays. Indicates the month index; S222: The calendar feature matrix and historical load sequence are concatenated through a feature concatenation layer to fuse physical history data and time label information. The concatenated data is then input into the Bi-LSTM model. The forward processing layer and backward processing layer of the Bi-LSTM model are used to extract the forward evolution pattern and reverse dependency features of the time series, respectively. The two are then fused to generate the final hidden layer state, as shown in formulas (13) and (14). (13) (14) In formula (13), For the Bi-LSTM model at time 1 The actual input, For historical load sequences, The calendar feature matrix; In formula (14), The predicted value of the power grid load t, These represent the weights and biases of the feature fusion layer, respectively.

[0008] In a preferred embodiment of the present invention, S2, the construction of a multimodal feature matrix using spatiotemporal coupling feature matrix, solar radiation prediction, power grid load prediction, structured static data, and external market data specifically includes the following steps: S231: A multimodal feature matrix is ​​constructed using the spatiotemporal coupling feature matrix, solar radiation forecast, power grid load forecast, structured static data, and external market data, as shown in formula (15). (15) In formula (15), For multimodal data feature matrix, The spatiotemporal coupling characteristic matrix, For a moment The predicted value of solar radiation, The predicted value of the power grid load t, For real-time time-of-use electricity pricing, For structured static data; S232: Standardize each component in the feature matrix of the multimodal data to eliminate dimensional differences between different variables, as shown in formulas (16) and (17). (16) (17) In formulas (16) and (17), For standardized components, The components of the feature matrix of multimodal data. and These are the mean and standard deviation of the historical data for the corresponding components. This is the feature matrix of multimodal data.

[0009] In a preferred embodiment of the present invention, in S3, a standard state space is constructed for the three independent agents using a multimodal data feature matrix. The three agents generate random actions through an actor network, and the probability density of the agent's selection of the corresponding action at the corresponding time step follows a Gaussian distribution random process to generate a random control strategy. Specifically, this includes the following steps: S311: Input the multimodal data feature matrix into the three independent agents, obtain the action mean and action standard deviation through the actor network, and then obtain the random actions generated by the three independent agents in the dimension of the multimodal feature matrix at the current time step through the action mean and action standard deviation. Then let the agent select the corresponding action at the corresponding time step according to the probability density of Gaussian distribution random process to generate random control strategy, as shown in formulas (18), (19), (20), (21), and (22). (18) (19) (20) (twenty one), (twenty two), In formulas (18), (19), (20), (21), (22), Represented as the mean of the action, and These are the weight matrix and bias vector of the action mean. For multimodal data feature matrix, For the standard deviation of the movement, These are the weight matrix and bias vector of the action standard deviation. yes The lower and upper bounds of a function. For random noise variables, For random actions, For the first An agent is in state Select action The probability density.

[0010] In a preferred embodiment of the present invention, in S3, the stochastic control policy is optimized using a proximal policy optimization algorithm through a shared actor network and an independent commentator network to obtain a deterministic control policy, and the deterministic control policy is output through a reinforcement learning controller, including the following steps: S321: The three intelligent agents are the first intelligent agent, the second intelligent agent, and the third intelligent agent. The first intelligent agent is used to minimize the economic operating cost of the distribution network, the second intelligent agent is used to smooth out grid fluctuations and reduce peak load impacts, and the third intelligent agent is used to minimize the charging and discharging switching frequency of the energy storage device and extend the life of the energy storage device. The corresponding objective functions are shown in formulas (23), (24), (25), and (26). (twenty three), (twenty four), (25) (26) In formulas (23), (24), (25), (26), As an immediate reward for the first intelligent agent, For a moment The power purchased by the power grid For a moment Real-time time-of-use electricity price This is the penalty coefficient for exceeding the limit. For a moment Peak load; For the immediate reward of the second intelligent agent, For visual risk weighting coefficients, The modulus length representing the fusion feature, For the target average load, Volatility weighting, For the past The standard deviation of net load over the period The standard deviation of the grid interaction power over the past T time steps; For the immediate reward of the third intelligent agent, To switch the penalty coefficient, For symbolic functions, For a moment The actual output power of the energy storage device For power amplitude weighting, For grid node loads, Real-time photovoltaic power generation. The energy storage power output by the intelligent agent; S322: Calculate the temporal difference error for each time step using the commentator network, as shown in formula (27). (27) In formula (27), For the first Temporal difference error of each agent For the first The instant reward corresponding to each intelligent agent. As a discount factor, For the first An independent estimate of the value of the current state by a network of commentators from individual agents. It is a multimodal feature matrix; S323: The generalized advantage estimate is calculated using the temporal difference error of the three agents, which guides the stochastic control strategy to generate an advantage value toward the target, as shown in formula (28). (28) In formula (28), for The dominance function value at time 1. As a discount factor, As a smoothing factor, For the first The first agent of the intelligent agent The timing difference error of the step; S324: Update the actor network parameters using the objective function of the proximal policy optimization algorithm, as shown in formulas (29), (30), and (31). (29) (30) (31), In formulas (29), (30), (31), The probability density ratio of the new and old strategies. Let be the probability density of the current policy. Let be the probability density of the previous strategy. For the strategy parameter set, for The dominance function value at time 1. for The dominance function value at time 1. This is a truncation function. For clipping hyperparameters; S325: Repeat S311 to generate new action mean, action standard deviation, and action; repeat S321-S324 to construct a multi-objective joint stability index. When the multi-objective joint stability index is less than a preset threshold, stop repeating and obtain the final strategy parameter set and the final action mean, as shown in formulas (32) and (33). (32), (33), In formulas (32) and (33), This is the current moving average reward. For the first An intelligent agent in The immediate reward obtained during each round of training, with 50 being the length of the sliding window. For the joint stability index of the system, ; S326: The final action mean is combined with the physical boundary of the energy storage device to generate a deterministic control strategy. The digital output through the neural network is then converted into actual control command output that meets the physical safety constraints by a reinforcement learning controller, as shown in formula (34). (34), In formula (34), The power to charge or discharge energy storage devices, when A value greater than zero indicates the energy storage device's intention to discharge. A value less than zero indicates the energy storage device's intention to charge. The percentage of rated power used to charge or discharge energy storage devices. This is the rated power.

[0011] In a preferred embodiment of the present invention, step S322 involves calculating the temporal difference error for each time step using a commentator network, a main generator, an auxiliary generator, and a discriminator, and includes the following steps: S3221: The main generator receives the grid node load, real-time photovoltaic power generation, and time-series monitoring dynamic data of a normally distributed random noise vector at the current time step. It concatenates the grid node load, real-time photovoltaic power generation, and normally distributed random noise vector in the same time step dimension, and generates a targeted short-term first disturbance sequence through neural network mapping to simulate prediction errors or sudden fluctuations, as shown in formula (35). (35), In formula (35), This is the first perturbation sequence. For grid node loads, Real-time photovoltaic power generation. Input is random noise. These are weight parameters; S3222: The auxiliary generator uses real historical error data to generate a reference sequence, calculates the statistical distribution difference between the first perturbation sequence and the reference sequence, and feeds it back to the main generator as a guiding loss, as shown in formula (36). (36) In formula (36), The loss value. This is the first perturbation sequence. For grid node loads, Real-time photovoltaic power generation. Input is random noise. These are weight parameters; S3223: The main generator and discriminator are jointly trained using a zero-sum game model. The discriminator is responsible for distinguishing between "true error data" and "first perturbation sequence," while the main generator continuously optimizes the first perturbation sequence and the guidance loss. To maximize the discrimination error, when the game reaches equilibrium, the output of the master generator is the second perturbation sequence, as shown in formula (37). (37) In formula (37), It is the second perturbation sequence. These are minima and maxima games, It is a binary classifier. For grid node loads, Real-time photovoltaic power generation. Input is random noise. This is the first perturbation sequence. These are the weighting coefficients. This is the loss value; S3224: The temporal difference error at each time step is calculated using the commentator network, as shown in formula (38). (38), In formula (38), For the first Temporal difference error of each agent For the first The instant reward corresponding to each intelligent agent. As a discount factor, For the first An independent estimate of the value of the current state by a network of commentators from individual agents. For multimodal feature matrices, It is the second perturbation sequence.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention preprocesses raw multidimensional operating condition data (such as structured static data, time-series monitoring dynamic data, external market data, and image data) to obtain a multimodal feature matrix. This multimodal feature matrix is ​​essentially a prediction of future environmental state data. It constructs a standard state space for the reinforcement learning agent, placing the agent in a complex environment. Then, three agents are established to generate a stochastic control strategy. Through a shared actor network and independent commentator networks, the proximal policy optimization (PPO) algorithm is used to optimize the stochastic control strategy, allowing it to be optimized under a set objective function. Simultaneously, through a set twin generator and discriminator, uncertain events that may occur in the future are predicted based on historical data. These uncertainties, i.e., the perturbation sequence, are used as reference conditions to further optimize the stochastic control strategy, ultimately yielding a deterministic control strategy. This allows the energy storage device to set its charging and discharging power according to the deterministic control strategy. Under complex environments and uncertain events, the photovoltaic-storage system adaptively schedules the charging and discharging power of the energy storage device based on the conditions of the photovoltaic system, energy storage device, and power grid system, improving the scheduling accuracy, stability, and adaptability of the photovoltaic-storage system. Attached Figure Description

[0013] Figure 1 This is a flowchart of the adaptive scheduling method for a photoelectric storage system based on multimodal data fusion, as described in this invention. Detailed Implementation

[0014] Typical embodiments embodying the features and advantages of the present invention will be specifically described in the following description. It should be understood that the present invention can have various variations in different embodiments without departing from the scope of the present invention, and the descriptions and illustrations herein are for illustrative purposes only and not intended to limit the present invention.

[0015] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0016] This embodiment discloses an adaptive scheduling method for a photoelectric storage system based on multimodal data fusion, such as... Figure 1 As shown, it includes the following steps: S1: Collect structured static data, time-series monitoring dynamic data, and external market data through corresponding sensors deployed in photovoltaic power generation systems, energy storage devices, and power grid systems, as well as image data of the photovoltaic panel itself and its surrounding physical environment through visual sensors. In S1, structured static data is used to define the system's boundary conditions, including the maximum charge and discharge power of the energy storage battery and the peak load limit allowed by the power grid. The time-series monitoring dynamic data includes light intensity, ambient temperature, relative humidity, real-time photovoltaic power generation, grid node load, and real-time state of charge of energy storage batteries; External market data includes real-time time-of-use electricity prices; A combination of a wide-angle lens and a fisheye lens is used to capture images of the photovoltaic panel itself and its surrounding physical environment.

[0017] In this embodiment, the collected data is transmitted to a central database or cloud server via industrial Ethernet or 5G network. The optical storage system uses a time-series database to store dynamic data and aligns it with a unified timestamp, providing a standardized data foundation for subsequent data fusion, feature extraction, and model training.

[0018] S2: The spatiotemporal coupling feature matrix and solar radiation prediction value are obtained from the time-series monitoring dynamic data and image data at each time step, respectively. At the same time, a calendar feature matrix is ​​constructed. A short-term load forecasting model based on Bi-LSTM is established using the time-series monitoring dynamic data and calendar feature matrix at each time step to obtain the power grid load forecast value. Then, a multimodal feature matrix is ​​constructed using the spatiotemporal coupling feature matrix, solar radiation prediction value, power grid load forecast value, structured static data, and external market data at each time step. In S2, the spatiotemporal coupling feature matrix and solar radiation prediction values ​​are obtained by monitoring dynamic data and image data over time, specifically including the following steps: S211: Preprocess the acquired images and divide them into... Each local feature is treated as a node in a graph structure. Edges are constructed based on the adjacency relationships of the features in physical space. A multi-layer graph attention network is used to iteratively update the node features. After the layers are aggregated, a set of feature vectors containing global spatial occlusion information is obtained, as shown in formulas (1) and (2). (1), (2), (3), In formulas (1), (2) and (3), For the first After layer attention aggregation, the updated result is the first layer. The new feature vector of each node This is the layer index of the graph attention network. It is a non-linear activation function used to enhance the non-linear expressive power of the model. For nodes neighborhood set, The attention coefficient represents the node. For nodes The influence weights, such as the influence weight of the shadow of nearby trees on photovoltaic panels, For the first The learnable linear transformation weight matrix of the layer, For the first Layer nodes eigenvectors, The spatial feature matrix is ​​the set of spatial feature matrices of all nodes, which fully describes the distribution of shading on the surface of the photovoltaic panel. In this embodiment, the image pixel relationship is transformed into a physical spatial relationship to quantify the degree of environmental occlusion.

[0019] In this embodiment, S211 calculates the association weights between nodes through an attention mechanism, enabling the model to automatically focus on key areas (i.e., high-weight neighborhoods) that are occluded by shadows or foreign objects, thus achieving a mapping from "pixel-level features" to "physical space occlusion semantics".

[0020] S212: The spatial feature matrix is ​​linearly transformed to obtain the query, key, and value matrix. The query, key, and value matrix is ​​then input into the ST-Transformer model. The standard self-attention method is used to output the spatiotemporal feature matrix with spatiotemporal dependence, as shown in formulas (4), (5), (6), and (7). (4), (5), (6), (7), In formulas (4), (5), (6), (7), These are query, key, and value matrices, respectively. For spatial characteristic matrix, These are the learnable linear mapping matrices, Let T be the spatiotemporal feature matrix, and T be the time step. The dimension of the key vector is denominator. It serves a scaling function to prevent the gradient from vanishing due to an excessively large dot product result; S213: The spatiotemporal feature matrix and the spatial feature matrix are fused along the channel dimension through a feature splicing layer to preserve local details and global temporal patterns, as shown in formula (8). (8), In formula (8), This is the characteristic matrix of environmental physical constraints. This is a global pooling operation. For spatial characteristic matrix, It is the spatiotemporal feature matrix; In this embodiment, the image data is subjected to deep processing through S211-S213 to extract potential physical constraint features.

[0021] S214: Extract the feature matrix of the time-series monitoring dynamic data, and fuse the feature matrix of the time-series monitoring dynamic data and the feature matrix of environmental physical constraints through a gating unit to obtain dual modal information of visual and numerical values, as shown in formula (9). (9), In formula (9), The spatiotemporal coupling characteristic matrix, These are adaptive gating weight coefficients, with values ​​ranging from 0 to 1. This is the characteristic matrix of environmental physical constraints. For time series feature extraction function, For time-series monitoring of dynamic data feature matrix; S215: Based on the spatiotemporal coupling feature matrix, the CNN-LSTM architecture is used to extract local spatiotemporal features, and a nonlinear autoregressive network is combined to capture the nonlinear relationship of the time series of solar radiation, as shown in formulas (10) and (11). (10) (11), In formula (10), For external input variables, For temperature, For humidity, For the current moment The previous Weather conditions at each time step, for index; In formula (11), For a moment The predicted value of solar radiation, and These are the mapping functions for nonlinear autoregressive networks and CNN-LSTM architectures, respectively. To delay the step size, These are measured values ​​of radiation at historical moments.

[0022] In this embodiment, , For real-time global radiation, This refers to theoretical outer space radiation.

[0023] In this embodiment, the nonlinear autoregressive network uses external meteorological factors as driving variables, which can effectively capture radiation nonlinear fluctuations caused by sudden weather changes and improve prediction accuracy.

[0024] In S2, obtaining the power grid load forecast value through time-series monitoring dynamic data and calendar feature matrix specifically includes the following steps: S221: Construct the calendar feature matrix to provide temporal context information for the Bi-LSTM model, as shown in formula (12). (12) In formula (12), This is a calendar feature matrix containing multi-dimensional time attributes. This represents the hourly index (range 0-23) of the current time point within a day, used to capture intraday peak and trough variations in load. A weekday type identifier (e.g., 1 for weekdays, 0 for weekends) is used to capture load differences caused by human work-life patterns. This indicates a statutory holiday (1 for holidays, 0 otherwise), used to correct load forecasting deviations during special holidays. This represents the month index (values ​​1-12), used to capture seasonal fluctuations in air conditioning or heating loads caused by temperature changes; S222: The calendar feature matrix and historical load sequence are concatenated through a feature concatenation layer to fuse physical history data and time label information. The concatenated data is then input into the Bi-LSTM model. The forward processing layer and backward processing layer of the Bi-LSTM model are used to extract the forward evolution pattern and reverse dependency features of the time series, respectively. The two are then fused to generate the final hidden layer state, as shown in formulas (13) and (14). (13) (14) In formula (13), For the Bi-LSTM model at time 1 The actual input, For historical load sequences, The calendar feature matrix; In formula (14), The predicted value of the power grid load t, These represent the weights and biases of the feature fusion layer, respectively.

[0025] In this embodiment, the historical load sequence is all historical load values ​​before time t.

[0026] In this embodiment, since the distribution network load is affected by user behavior and has significant daily and weekly periodicity, a short-term load forecasting model based on Bi-LSTM is established to capture the time-series periodicity characteristics of user-side load.

[0027] In S2, the construction of a multimodal feature matrix using spatiotemporal coupling feature matrix, solar radiation forecast, power grid load forecast, structured static data, and external market data specifically includes the following steps: S231: A multimodal feature matrix is ​​constructed using the spatiotemporal coupling feature matrix, solar radiation forecast, power grid load forecast, structured static data, and external market data, as shown in formula (15). (15) In formula (15), For multimodal data feature matrix, The spatiotemporal coupling characteristic matrix, For a moment The predicted value of solar radiation, The predicted value of the power grid load t, For real-time time-of-use electricity pricing, For structured static data; S232: Standardize each component in the feature matrix of the multimodal data to eliminate dimensional differences between different variables, as shown in formulas (16) and (17). (16) (17) In formulas (16) and (17), For standardized components, The components of the feature matrix of multimodal data. and These are the mean and standard deviation of the historical data for the corresponding components. This is the feature matrix of multimodal data.

[0028] S3: Construct a standard state space for three independent agents using the multimodal data feature matrix. The three agents generate random actions through an actor network and allow the agents to select actions at corresponding time steps with a probability density that follows a Gaussian distribution. Generate a stochastic control policy. Optimize the stochastic control policy using a proximal policy optimization algorithm through a shared actor network and independent commentator networks to obtain a deterministic control policy. Output the deterministic control policy through a reinforcement learning controller. In S3, a standard state space is constructed for three independent agents using multimodal data feature matrices. The three agents generate random actions through an actor network, and the probability density of the agent's choice of action at each time step follows a Gaussian distribution, generating a stochastic control policy. The specific steps include: S311: Input the multimodal data feature matrix into the three independent agents, obtain the action mean and action standard deviation through the actor network, and then obtain the random actions generated by the three independent agents in the dimension of the multimodal feature matrix at the current time step through the action mean and action standard deviation. Then let the agent select the corresponding action at the corresponding time step according to the probability density of Gaussian distribution random process to generate random control strategy, as shown in formulas (18), (19), (20), (21), and (22). (18) (19) (20) (twenty one), (twenty two), In formulas (18), (19), (20), (21), (22), Represented as the mean of the action, and These are the weight matrix and bias vector of the action mean. For multimodal data feature matrix, For the standard deviation of the movement, These are the weight matrix and bias vector of the action standard deviation. yes The lower and upper bounds of a function. For random noise variables, For random actions, For the first An agent is in state Select action The probability density.

[0029] In this embodiment, the intelligent agent refers to a virtual decision-making unit constructed based on a deep reinforcement learning algorithm. Specifically, this system establishes three functionally heterogeneous independent intelligent agents that share the same perception layer. However, it possesses an independent value assessment system. The first agent focuses on economic optimization, the second on grid stability, and the third on extending equipment lifespan. Through collaborative training, the three agents jointly adjust the policy network parameters, ultimately outputting control commands that balance multi-dimensional optimization objectives.

[0030] In S3, a stochastic control policy is optimized using a proximal policy optimization algorithm through a shared actor network and an independent commentator network to obtain a deterministic control policy. The deterministic control policy is then output through a reinforcement learning controller, including the following steps: S321: The three intelligent agents are the first intelligent agent, the second intelligent agent, and the third intelligent agent. The first intelligent agent is used to minimize the economic operating cost of the distribution network, the second intelligent agent is used to smooth out grid fluctuations and reduce peak load impacts, and the third intelligent agent is used to minimize the charging and discharging switching frequency of the energy storage device and extend the life of the energy storage device. The corresponding objective functions are shown in formulas (23), (24), (25), and (26). (twenty three), (twenty four), (25) (26) In formulas (23), (24), (25), (26), The immediate reward for the first agent (the immediate reward refers to the single-step reward fed back by the environment after the agent performs an action at time t). For a moment The power purchased by the power grid For a moment Real-time time-of-use electricity price This is the penalty coefficient for exceeding the limit. For a moment Peak load; For the immediate reward of the second intelligent agent, For visual risk weighting coefficients, The modulus length represents the fusion characteristic (representing the severity of the weather). For the target average load, Volatility weighting, For the past The standard deviation of net load over the period For the past The standard deviation of the power grid interaction power within a time step is used to quantify the severity of fluctuations. For the immediate reward of the third intelligent agent, To switch the penalty coefficient and prevent the battery from oscillating at high frequencies between charging and discharging, This is a sign function (for determining the charging and discharging direction). For a moment The actual output power of the energy storage device Power amplitude weighting is used to prevent damage to battery life from high-rate charging and discharging. For grid node loads, Real-time photovoltaic power generation. The energy storage power output by the intelligent agent; In this embodiment, the objective function of the first intelligent agent is used to drive the first intelligent agent to utilize energy storage for "low storage and high generation" to achieve arbitrage and reduce demand electricity costs.

[0031] In this embodiment, Index of the current time step. The length of the historical sliding window (e.g., the number of time steps in the past hour).

[0032] In this embodiment, multiple agents can independently evaluate the merits of their current actions based on different objectives, without being interfered with by the objectives of other agents.

[0033] S322: Calculate the temporal difference error for each time step using the commentator network, as shown in formula (27). (27) In formula (27), For the first Temporal difference error of each agent For the first The instant reward corresponding to each intelligent agent. This is a discount factor, representing the decay of future rewards. For the first The independent estimates of the value of the current state by the commentator network of each agent are typically a scalar value output by a multi-layer fully connected MLP after weighted activation. It is a multimodal feature matrix; In this embodiment, step S322 involves calculating the temporal difference error for each time step using a commentator network, a main generator, an auxiliary generator, and a discriminator. This includes the following steps: S3221: The main generator receives the grid node load, real-time photovoltaic power generation, and time-series monitoring dynamic data of a normally distributed random noise vector at the current time step. It concatenates the grid node load, real-time photovoltaic power generation, and normally distributed random noise vector in the same time step dimension, and generates a targeted short-term first disturbance sequence through neural network mapping to simulate prediction errors or sudden fluctuations, as shown in formula (35). (35), In formula (35), This is the first perturbation sequence. For grid node loads, Real-time photovoltaic power generation. Input is random noise. These are weight parameters; S3222: The auxiliary generator uses real historical error data to generate a reference sequence, calculates the statistical distribution difference between the first perturbation sequence and the reference sequence, and feeds it back to the main generator as a guiding loss, as shown in formula (36). (36) In formula (36), The loss value. This is the first perturbation sequence. For grid node loads, Real-time photovoltaic power generation. Input is random noise. These are weight parameters; S3223: The main generator and discriminator are jointly trained using a zero-sum game model. The discriminator is responsible for distinguishing between "true error data" and "first perturbation sequence," while the main generator continuously optimizes the first perturbation sequence and the guidance loss. To maximize the discrimination error, when the game reaches equilibrium, the output of the master generator is the second perturbation sequence, as shown in formula (37). (37) In formula (37), It is the second perturbation sequence. These are a minimax game, where the discriminator D attempts to maximize the value in parentheses, and the master generator... Try to minimize the value inside the parentheses. It is a binary classifier that outputs a probability value between 0 and 1. The closer the value is to 1, the more likely the discriminator is to consider the input data to be true. For grid node loads, Real-time photovoltaic power generation. Input is random noise. This is the first perturbation sequence. This is the weighting coefficient, typically taken as 0.5. This is the loss value; S3224: The temporal difference error at each time step is calculated using the commentator network, as shown in formula (38). (38), In formula (38), For the first Temporal difference error of each agent For the first The instant reward corresponding to each intelligent agent. As a discount factor, For the first An independent estimate of the value of the current state by a network of commentators from individual agents. For multimodal feature matrices, It is the second perturbation sequence.

[0034] In this embodiment, This indicates that the extreme perturbation data generated in the adversarial state will be superimposed onto the real-collected multimodal feature matrix.

[0035] S323: The generalized advantage estimate is calculated using the temporal difference error of the three agents, which guides the stochastic control strategy to generate an advantage value toward the target, as shown in formula (28). (28) In formula (28), for The dominance function value at time 1. As a discount factor, This is the smoothing factor (usually taken as 0.95). For the first The first agent of the intelligent agent The timing difference error of the step; In this embodiment, if A positive value indicates that the action is beneficial in the long run, and its probability should be increased.

[0036] S324: Update the actor network parameters using the objective function of the proximal policy optimization algorithm, as shown in formulas (29), (30), and (31). (29) (30) (31), In formulas (29), (30), (31), The probability density ratio of the new and old strategies. This represents the probability density that is currently being updated via gradient descent. The probability density used during the trajectory collection phase (when sampling data). For the strategy parameter set, The function forces gradient truncation, thus ensuring the monotonical non-decreasing nature and stability of training in complex power grid environments. for The dominance function value at time 1. for The dominance function value at time 1. For truncation function, when r < Output via function ,when , , , This is a pruning hyperparameter used to prevent the strategy from resembling amplitude; it is typically set to 0.2. S325: Repeat S311 to generate new action mean, action standard deviation, and action; repeat S321-S324 to construct a multi-objective joint stability index. When the multi-objective joint stability index is less than a preset threshold, stop repeating and obtain the final strategy parameter set and the final action mean, as shown in formulas (32) and (33). (32), (33), In formulas (32) and (33), This is the current moving average reward. For the first An intelligent agent in The immediate reward obtained during each round of training, with 50 being the length of the sliding window. For the joint stability index of the system, It is a local constant to prevent the denominator from being zero; S326: The final action mean is combined with the physical boundary of the energy storage device to generate a deterministic control strategy. The digital output through the neural network is then converted into actual control command output that meets the physical safety constraints by a reinforcement learning controller, as shown in formula (34). (34), In formula (34), The power to charge or discharge energy storage devices, when A value greater than zero indicates the energy storage device's intention to discharge. A value less than zero indicates the energy storage device's intention to charge. The percentage of rated power used to charge or discharge energy storage devices. This is the rated power.

[0037] In this embodiment, the value is the average of the actions of the three agents. The neural network internally processes unitless mathematical values, and the range may be... Without intervention, the energy storage device will not know how much it needs to charge and discharge. This can be determined by multiplying by the rated power. This instantly transforms dimensionless numbers into physical quantities with units (kW). For example... If the output is 0.5 and the battery rating is 100kW, then the instruction is... .

[0038] In this embodiment, Ensure the output is between [-1, 1].

[0039] S4: Repeat S1-S3, and the optical storage system stably outputs the corresponding deterministic control strategy at each time step.

[0040] In this embodiment, continuous time is divided into segments, each segment being a time step. In this embodiment, 10 minutes is used as a time step.

[0041] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for adaptive dispatching of optical storage systems based on multi-modal data fusion, characterized in that: Includes the following steps: S1: Collect structured static data, time-series monitoring dynamic data, and external market data through corresponding sensors deployed in photovoltaic power generation systems, energy storage devices, and power grid systems, as well as image data of the photovoltaic panel itself and its surrounding physical environment through visual sensors; S2: The spatiotemporal coupling feature matrix and solar radiation prediction value are obtained from the time-series monitoring dynamic data and image data at each time step, respectively. At the same time, a calendar feature matrix is ​​constructed. A short-term load forecasting model based on Bi-LSTM is established using the time-series monitoring dynamic data and calendar feature matrix at each time step to obtain the power grid load forecast value. Then, a multimodal feature matrix is ​​constructed using the spatiotemporal coupling feature matrix, solar radiation prediction value, power grid load forecast value, structured static data, and external market data at each time step. S3: Construct a standard state space for three independent agents using the multimodal data feature matrix. The three agents generate random actions through an actor network and allow the agents to select actions at corresponding time steps with a probability density that follows a Gaussian distribution. Generate a stochastic control policy. Optimize the stochastic control policy using a proximal policy optimization algorithm through a shared actor network and independent commentator networks to obtain a deterministic control policy. Output the deterministic control policy through a reinforcement learning controller. S4: Repeat S1-S3, and the optical storage system stably outputs the corresponding deterministic control strategy at each time step.

2. The adaptive scheduling method for a photoelectric storage system based on multimodal data fusion according to claim 1, characterized in that: In S1, structured static data is used to define the system's boundary conditions, including the maximum charge and discharge power of the energy storage battery and the peak load limit allowed by the power grid. The time-series monitoring dynamic data includes light intensity, ambient temperature, relative humidity, real-time photovoltaic power generation, grid node load, and real-time state of charge of energy storage batteries; External market data includes real-time time-of-use electricity prices; A combination of a wide-angle lens and a fisheye lens is used to capture images of the photovoltaic panel itself and its surrounding physical environment.

3. The multi-modal data fusion based adaptive scheduling method for optical storage system according to claim 1, wherein: In S2, the spatiotemporal coupling feature matrix and solar radiation prediction values ​​are obtained by monitoring dynamic data and image data over time, specifically including the following steps: S211: Preprocess the acquired images and divide them into... Each local feature is treated as a node in a graph structure. Edges are constructed based on the adjacency relationships of the features in physical space. A multi-layer graph attention network is used to iteratively update the node features. After the layers are aggregated, a set of feature vectors containing global spatial occlusion information is obtained, as shown in formulas (1) and (2). （1）， （2）， （3）， In formulas (1), (2) and (3), For the first After layer attention aggregation, the updated result is the 1st generation. The new feature vector of each node This is the layer index of the graph attention network. It is a non-linear activation function. For nodes neighborhood set, Attention coefficient For the first The learnable linear transformation weight matrix of the layer, For the first Layer nodes eigenvectors, It is the spatial characteristic matrix; S212: The spatial feature matrix is ​​linearly transformed to obtain the query, key, and value matrix. The query, key, and value matrix is ​​then input into the ST-Transformer model. The standard self-attention method is used to output the spatiotemporal feature matrix with spatiotemporal dependence, as shown in formulas (4), (5), (6), and (7). （4）， （5）， （6）， （7）， In formulas (4), (5), (6), (7), These are query, key, and value matrices, respectively. For spatial characteristic matrix, These are the learnable linear mapping matrices, Let T be the spatiotemporal feature matrix, and T be the time step. The dimension of the key vector; S213: The spatiotemporal feature matrix and the spatial feature matrix are fused along the channel dimension through a feature splicing layer, preserving local details and global temporal patterns, as shown in formula (8). （8）， In formula (8), This is the characteristic matrix of environmental physical constraints. This is a global pooling operation. For spatial characteristic matrix, It is the spatiotemporal feature matrix; S214: Extract the feature matrix of the time-series monitoring dynamic data, and fuse the feature matrix of the time-series monitoring dynamic data and the feature matrix of environmental physical constraints through a gating unit to obtain dual modal information of visual and numerical values, as shown in formula (9). （9）， In formula (9), The spatiotemporal coupling characteristic matrix, For adaptive gating weight coefficients, This is the characteristic matrix of environmental physical constraints. For time series feature extraction function, For time-series monitoring of dynamic data feature matrix; S215: Based on the spatiotemporal coupling feature matrix, the CNN-LSTM architecture is used to extract local spatiotemporal features, and a nonlinear autoregressive network is combined to capture the nonlinear relationship of the time series of solar radiation, as shown in formulas (10) and (11). （10）， （11）， In formula (10), For external input variables, For temperature, For humidity, For the current moment The previous Weather conditions at each time step, for index; In formula (11), For a moment The predicted value of solar radiation, and These are the mapping functions for nonlinear autoregressive networks and CNN-LSTM architectures, respectively. To delay the step size, These are measured values ​​of radiation at historical moments.

4. The multi-modal data fusion based adaptive scheduling method for optical storage system according to claim 3, wherein: In S2, obtaining the power grid load forecast value through time-series monitoring dynamic data and calendar feature matrix specifically includes the following steps: S221: Construct the calendar feature matrix to provide temporal context information for the Bi-LSTM model, as shown in formula (12). （12）， In formula (12), The calendar feature matrix, This represents the hourly index of the current time within a day. Workday type identifier Indicates the sign for statutory holidays. Indicates the month index; S222: The calendar feature matrix and historical load sequence are concatenated through a feature concatenation layer to fuse physical history data and time label information. The concatenated data is then input into the Bi-LSTM model. The forward processing layer and backward processing layer of the Bi-LSTM model are used to extract the forward evolution pattern and reverse dependency features of the time series, respectively. The two are then fused to generate the final hidden layer state, as shown in formulas (13) and (14). （13）， （14）， In formula (13), For the Bi-LSTM model at time 10:00 The actual input, For historical load sequences, The calendar feature matrix; In equation (14), a predicted value of the grid load of t, are the weight and bias of the feature fusion layer, respectively.

5. The multi-modal data fusion based adaptive scheduling method for optical storage system according to claim 4, wherein: In S2, the construction of a multimodal feature matrix using spatiotemporal coupling feature matrix, solar radiation forecast, power grid load forecast, structured static data, and external market data specifically includes the following steps: S231: A multimodal feature matrix is ​​constructed using the spatiotemporal coupling feature matrix, solar radiation forecast, power grid load forecast, structured static data, and external market data, as shown in formula (15). （15）， In formula (15), For multimodal data feature matrix, The spatiotemporal coupling characteristic matrix, For a moment The predicted value of solar radiation, The predicted value of the power grid load t, For real-time time-of-use electricity pricing, For structured static data; S232: Standardize each component in the feature matrix of the multimodal data to eliminate dimensional differences between different variables, as shown in formulas (16) and (17). （16）， （17）， In formulas (16) and (17), For standardized components, The components of the feature matrix of multimodal data. and These are the mean and standard deviation of the historical data for the corresponding components. This is the feature matrix of multimodal data.

6. The multi-modal data fusion based adaptive scheduling method for optical storage system according to claim 1, wherein: In S3, a standard state space is constructed for three independent agents using multimodal data feature matrices. The three agents generate random actions through an actor network, and the probability density of the agent's choice of action at each time step follows a Gaussian distribution, generating a stochastic control policy. The specific steps include: S311: Input the multimodal data feature matrix into the three independent agents, obtain the action mean and action standard deviation through the actor network, and then obtain the random actions generated by the three independent agents in the dimension of the multimodal feature matrix at the current time step through the action mean and action standard deviation. Then let the agent select the corresponding action at the corresponding time step according to the probability density of Gaussian distribution random process to generate random control strategy, as shown in formulas (18), (19), (20), (21), and (22). （18）， （19）， （20）， （21）， （22）， In formulas (18), (19), (20), (21), (22), Represented as the mean of the action, and These are the weight matrix and bias vector of the action mean. For multimodal data feature matrix, For the standard deviation of the movement, These are the weight matrix and bias vector of the action standard deviation. yes The lower and upper bounds of a function. For random noise variables, For random actions, For the first An agent is in state Select action The probability density.

7. The multi-modal data fusion based adaptive scheduling method for optical storage system according to claim 6, characterized in that: In S3, a stochastic control policy is optimized using a proximal policy optimization algorithm through a shared actor network and an independent commentator network to obtain a deterministic control policy. The deterministic control policy is then output through a reinforcement learning controller, including the following steps: S321: The three intelligent agents are the first intelligent agent, the second intelligent agent, and the third intelligent agent. The first intelligent agent is used to minimize the economic operating cost of the distribution network, the second intelligent agent is used to smooth out grid fluctuations and reduce peak load impacts, and the third intelligent agent is used to minimize the charging and discharging switching frequency of the energy storage device and extend the life of the energy storage device. The corresponding objective functions are shown in formulas (23), (24), (25), and (26). （23）， （24）， （25）， （26）， In formulas (23), (24), (25), and (26), is the instantaneous reward of the first agent, is the time is the grid purchase power, is the time is the real-time time-of-use price of electricity, is the out-of-limit penalty coefficient, is the peak load at time is the peak load at time instantaneous reward for the second agent, visual risk weight coefficient, modulus of the fused feature, target average load, volatility weight, past standard deviation of the net load in the time period, standard deviation of the grid interaction power in the past T time steps; an immediate reward for the third agent, a switching penalty coefficient, a sign function, a time instant an actual output power of the energy storage device, a power amplitude weight, a grid node load, a real-time photovoltaic power generation, a storage power output by the agent; S322: Calculate the temporal difference error for each time step using the commentator network, as shown in formula (27). （27）， In equation (27), is the time difference error of the th agent, is the immediate reward corresponding to the th agent, is the discount factor, is the critic network of the th agent's independent estimate of the current state value, is the multi-modal feature matrix; S323: The generalized advantage estimate is calculated using the temporal difference error of the three agents, which guides the stochastic control strategy to generate an advantage value toward the target, as shown in formula (28). （28）， In Equation (28), is the advantage function value at time step t, is a discount factor, is a smoothing factor, is the time step t for the th agent. S324: Update the actor network parameters using the objective function of the proximal policy optimization algorithm, as shown in formulas (29), (30), and (31). （29）， （30）， （31）， In formulas (29), (30), (31), The probability density ratio of the new and old strategies. Let be the probability density of the current policy. Let be the probability density of the previous strategy. For the strategy parameter set, for The dominance function value at time 1. for The dominance function value at time 1. This is a truncation function. For clipping hyperparameters; S325: Repeat S311 to generate new action mean, action standard deviation, and action; repeat S321-S324 to construct a multi-objective joint stability index. When the multi-objective joint stability index is less than a preset threshold, stop repeating and obtain the final strategy parameter set and final action mean, as shown in formulas (32) and (33). （32）， （33）， In formulas (32) and (33), This is the current moving average reward. For the first An intelligent agent in The immediate reward obtained during each round of training, with 50 being the length of the sliding window. For the joint stability index of the system, ; S326: The final action mean is combined with the physical boundary of the energy storage device to generate a deterministic control strategy. The digital output through the neural network is then converted into actual control command output that meets the physical safety constraints by a reinforcement learning controller, as shown in formula (34). （34）， In equation (34), is the charging or discharging power of the energy storage device, when is greater than zero, indicates the energy storage device discharging intention, when is less than zero, indicates the energy storage device charging intention, is the charging or discharging power of the energy storage device as a percentage of the rated power, is the rated power.

8. The multi-modal data fusion based adaptive scheduling method for optical storage system according to claim 7, wherein: In S322, the temporal difference error at each time step is calculated through a commentator network, a main generator, an auxiliary generator, and a discriminator, including the following steps: S3221: The main generator receives the grid node load, real-time photovoltaic power generation, and time-series monitoring dynamic data of a normally distributed random noise vector at the current time step. It concatenates the grid node load, real-time photovoltaic power generation, and normally distributed random noise vector in the same time step dimension, and generates a targeted short-term first disturbance sequence through neural network mapping to simulate prediction errors or sudden fluctuations, as shown in formula (35). （35）， In equation (35), is the first disturbance sequence, is the grid node load, is the real-time PV generation power, is the random noise input, is the weight parameter; S3222: The auxiliary generator uses real historical error data to generate a reference sequence, calculates the statistical distribution difference between the first perturbation sequence and the reference sequence, and feeds it back to the main generator as a guiding loss, as shown in formula (36). （36）， In equation (36), is the loss value, is the first disturbance sequence, is the grid node load, is the real-time PV power generation, is the random noise input, is the weight parameter; S3223: the main generator and the discriminator are jointly trained in a zero-sum game mode, the discriminator is responsible for distinguishing "real error data" and "first perturbation sequence", the main generator continuously optimizes the first perturbation sequence and the guidance loss , to maximize the discrimination error, when the game reaches equilibrium, the output of the main generator is the second perturbation sequence, as shown in equation (37), （37）， In equation (37), is a second disturbance sequence, are minimax games, respectively, is a binary classification discriminator, is a grid node load, is a real-time PV generation power, is a random noise input, is a first disturbance sequence, is a weight coefficient, is a loss value; S3224: The temporal difference error at each time step is calculated using the commentator network, as shown in formula (38). （38）， In Equation (38), is the time difference error for the th agent, is the immediate reward corresponding to the th agent, is the discount factor, is the critic network's independent estimate of the current state value for the th agent, is the multi-modal feature matrix, is the second perturbation sequence.

Images