Load scheduling method for new energy station

Abstract

The present application relates to the technical field of power system operation control, in particular to a load scheduling method for new energy station, which collects multi-dimensional operation data through the sensing unit deployed in the station. Based on these data, a load prediction model is constructed to generate short-term and ultra-short-term charging load prediction results. Based on the prediction results, an improved multi-agent deep deterministic policy gradient algorithm is used to construct a scheduling optimization framework in the decision core. According to the real-time updated electricity price, transformer capacity limit, vehicle reservation time and user charging preference parameters, the algorithm calculates a set of dynamic constraint conditions. The scheduling optimization framework solves under this dynamic constraint to roll out the global optimal charging power allocation strategy within a future period. Finally, the strategy is decomposed into individual power control instructions for each charging pile, and combined with the vehicle battery state and user preference to generate an orderly charging guidance scheme for specific electric vehicle users.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, and in particular to a load dispatching method for new energy power plants. Background Technology

[0002] In the process of building new power systems, the integration of numerous electric vehicle charging facilities in renewable energy power plants has become a typical scenario. These plants must simultaneously address the intermittency of renewable energy output, the volatility of grid-side time-of-use pricing, and the randomness of user charging demand to achieve economical, reliable, and user-friendly grid-connected operation. Existing scheduling technologies for charging loads largely rely on statistical analysis of historical aggregated load curves or start-stop control based on fixed rules. These methods typically treat all charging loads within the power plant as a whole or a few homogeneous clusters, with relatively singular data dimensions for decision-making, often considering only grid pricing or simple load thresholds. Due to the lack of refined perception and integration of multi-dimensional information such as the battery status of individual electric vehicles within the power plant, user-specific charging time windows, and power preferences, traditional methods struggle to characterize and respond to real-time dynamic differences in charging demand. Regarding optimization models, conventional centralized optimization or rule-based methods face challenges in computational efficiency and solution quality when dealing with large-scale, high-dimensional continuous decision variables and complex time-varying constraints. When grid electricity price signals are updated, vehicles connect in real time, or users temporarily modify their charging plans, the preset static optimization models or fixed strategies are difficult to adapt quickly, resulting in insufficient economic efficiency of the dispatching scheme or violation of equipment safety constraints. In addition, most dispatching strategies focus on the overall benefits of the charging station, and their output is a total power command. They lack a mechanism to transparently and reasonably decompose this command to each specific charging pile and interact with users, which reduces user participation and the feasibility of the scheme. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies and propose a load dispatching method for new energy power plants.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a load dispatching method for new energy power plants, comprising:

[0005] The sensing units deployed at new energy power stations collect multi-dimensional operational data, including the real-time status of charging piles, battery status information of connected electric vehicles, time-of-use electricity price signals of the power grid, and real-time load data of the power station's power supply area.

[0006] Based on the multi-dimensional operational data, a load forecasting model is constructed and run to generate short-term and ultra-short-term charging load forecasts for new energy power stations in future periods.

[0007] Based on the short-term and ultra-short-term charging load forecast results, an improved multi-agent deep deterministic strategy gradient algorithm is used in the decision core to construct a scheduling optimization framework with the optimization objectives of reducing station operating costs, smoothing grid load fluctuations, and meeting user charging needs.

[0008] The improved multi-agent deep deterministic strategy gradient algorithm calculates a dynamic set of constraints based on the real-time updated time-of-use electricity price signal, the capacity limit of the main transformer at the power station, the scheduled charging time of the vehicle, and the charging preference parameters set by the user.

[0009] The scheduling optimization framework solves the problem under the dynamic set of constraints and outputs the globally optimal charging power allocation strategy for the next optimization cycle.

[0010] The global optimal charging power allocation strategy is decomposed into individual power control commands for each charging pile, and an orderly charging guidance scheme for specific electric vehicle users is generated based on vehicle battery status information and user charging preference parameters.

[0011] As a further aspect of the present invention, a load forecasting model is constructed and run based on the multi-dimensional operational data, including:

[0012] The multi-dimensional operational data is preprocessed, including removing outlier data points, filling in missing data values, and aligning data at different time scales to time series.

[0013] Charging load-related features are extracted from the preprocessed data. These features include historical load curves for the same period, the number of vehicles charging in real time, the average battery state of charge of the vehicles, and the spatiotemporal distribution information of future scheduled charging orders.

[0014] The features are input into a short-term prediction network, which uses a convolutional neural network combined with a gated recurrent unit structure. The network uses hours as the time step to predict the total charging load curve of new energy power stations for the next 24 hours and generates the short-term charging load prediction result.

[0015] The features are input into an ultra-short-term prediction network, which adopts a temporal convolutional network structure and uses minutes as the time step to predict the total charging load curve of new energy power stations for the next fifteen minutes to one hour, thereby generating the ultra-short-term charging load prediction result.

[0016] As a further aspect of the present invention, the working steps of the improved multi-agent deep deterministic policy gradient algorithm are as follows:

[0017] Each charging pile in a new energy power station is defined as an intelligent agent, and an independent execution network and evaluation network are built for it. All intelligent agents share a global environmental state, which is composed of the short-term charging load prediction results, the ultra-short-term charging load prediction results, and the time-of-use electricity price signal.

[0018] Each agent's execution network outputs the suggested charging power of its corresponding charging station in the next moment based on the current global environment state. The suggested charging power of all agents together constitutes a joint action.

[0019] The joint action is input into a centralized network of critics, which evaluates the long-term expected total reward of the joint action based on the global environment state and the joint action, and outputs a global state-action value function.

[0020] Each agent's evaluation network uses the global state-action value function and combines it with the agent's local observation information to calculate the quality of the agent's actions, which is used to guide the parameter updates of its execution network.

[0021] At each training time step, the agent samples a batch of historical experience data from the experience replay pool, which includes historical environmental states, historical joint actions, immediate rewards obtained, and new environmental states.

[0022] Using the sampled historical experience data, the parameters of the network of critics in the set are updated by minimizing the temporal difference error;

[0023] The parameters of the execution network for each agent are updated via a policy gradient method in the direction that improves the global state-action value function evaluated by the commentator network in the set.

[0024] As a further aspect of the present invention, the evaluation network for each agent utilizes the global state-action value function and combines it with the agent's local observation information to calculate the merits of the agent's actions, which is used to guide the parameter updates of its execution network, including:

[0025] The evaluation network receives the global state-action value function output from the network of critics in the set, and uses it as a global benchmark reference value for evaluating the actions of the agent.

[0026] Meanwhile, the evaluation network receives local observation information from the agent, including the real-time occupancy status of the corresponding charging pile, the battery charge status of the currently serving vehicle, the charging time of the vehicle, and the user's charging preference parameters.

[0027] The global state-action value function is fused with the local observation information of the agent, and then input into the multi-layer neural network of the evaluation network.

[0028] The evaluation network is computed through forward propagation of a multi-layer neural network to output a local state-action value function estimate. The state-action value function estimate represents the long-term value of executing the actions suggested by the network under the current global environment state and the local observations of the agent.

[0029] The difference between the estimated state-action value function and the actual immediate reward plus the estimated value of the next state is calculated to obtain the temporal difference error for this agent.

[0030] The network parameters of the evaluation network are updated using the temporal difference error through the gradient backpropagation algorithm to reduce its error in action value assessment.

[0031] Based on the updated evaluation network, the value gradient of the action suggested by the execution network is calculated, and the value gradient indicates how the action should be adjusted to improve long-term returns;

[0032] The value gradient is passed to the execution network, which then uses a policy gradient ascent method to adjust its network parameters based on this gradient, making the execution network more inclined to output actions that can obtain higher evaluations.

[0033] As a further aspect of the present invention, the design of the immediate reward is determined based on the optimization objective and the set of dynamic constraints, including:

[0034] After the joint action is executed, the total electricity cost of the new energy power station in a certain period of time is calculated based on the time-of-use electricity price signal and the total charging power, and a cost reward component that is negatively correlated with the total electricity cost is generated.

[0035] Calculate the smoothness index of the total charging power curve of new energy power stations. The smoothness index is the sum of squares of the rate of change of total charging power within a time period, and generate a smoothing reward component that is negatively correlated with the smoothness index.

[0036] The system calculates the ratio of the actual charging amount of all charging vehicles to the user's expected charging demand within a given time period, and generates a demand fulfillment reward component that is positively correlated with the fulfillment ratio.

[0037] Check whether the combined action violates the set of dynamic constraints. Apply a large negative penalty to actions that violate the transformer capacity limit, and apply a penalty that increases with time urgency to actions that violate the user-set charging cutoff time constraint.

[0038] The cost reward component, smoothing reward component, demand satisfaction reward component, and various penalties are weighted and summed. The result is used as the instantaneous reward obtained by all agents at the time step and fed back to the improved multi-agent deep deterministic policy gradient algorithm.

[0039] As a further aspect of the present invention, the improved multi-agent deep deterministic policy gradient algorithm calculates a dynamic set of constraints based on the real-time updated time-of-use electricity price signal, the capacity limit of the main transformer at the power station, the vehicle's scheduled charging time, and the user-set charging preference parameters, including:

[0040] The time-of-use electricity price signal is acquired in real time, and the electricity price change time points within the future optimization cycle are determined according to the electricity price time period division boundary. Based on the electricity price change time points, power cost weight vectors for different time periods are generated.

[0041] The current load rate and rated capacity of the main transformer in the power station are obtained in real time. The real-time available capacity of the transformer at each moment in the optimization cycle is calculated. The changing trend of capacity limit is predicted by combining the historical load curve and a time-indexed capacity limit sequence is generated.

[0042] Iterate through all connected and reserved vehicles, extract the reserved charging deadline and current time for each vehicle, calculate the remaining available time window for each vehicle's charging needs, and combine the vehicle's current battery state of charge and target battery state of charge to generate the minimum average charging power requirement for each vehicle.

[0043] The system analyzes the user's charging preference parameters, extracts the user's acceptable latest completion time, expected range of battery capacity increase, and sensitivity level to charging costs, compares the latest completion time with the scheduled charging deadline, and takes the stricter time as the user's effective charging deadline constraint.

[0044] By integrating the power cost weight vector, the capacity upper limit sequence, the minimum average charging power requirement of each vehicle, and the effective charging cutoff time constraint of each user, a dynamic constraint mathematical expression containing inequality constraints and equality constraints is constructed. The dynamic constraint set is composed of the real-time calculation results of the dynamic constraint mathematical expression.

[0045] As a further aspect of the present invention, the scheduling optimization framework solves the problem under the dynamic set of constraints, including:

[0046] At the beginning of each optimization cycle, the state of all agents in the improved multi-agent deep deterministic policy gradient algorithm is initialized. The state includes the current real-time state of each charging pile, the battery status information of each vehicle, the current electricity price period, and the future short-term charging load prediction results.

[0047] The improved multi-agent deep deterministic policy gradient algorithm aims to maximize long-term cumulative rewards by exploring in the continuous action space and simulating charging power allocation scenarios for multiple future time steps.

[0048] During the simulation, at each time step forward, the current virtual joint action needs to be substituted into the dynamic set of constraints for feasibility verification.

[0049] If the verification passes, the virtual reward is calculated according to the reward function, the agent state is updated, and the simulation continues.

[0050] If the verification fails, a constraint penalty is imposed on the joint action, and the agent is guided to avoid such actions in subsequent simulations. At the same time, local replanning may be triggered.

[0051] After completing a preset number of simulated exploration steps, the action sequence that obtains the highest cumulative reward is selected as the optimization result of the current optimization cycle, and the joint action of the first time step of the action sequence is output as the initial part of the global optimal charging power allocation strategy.

[0052] As a further aspect of the present invention, the feasibility verification process specifically includes:

[0053] Extract the virtual joint action at the current time step, i.e., the virtual charging power suggestion value of all charging piles, and sum them to obtain the total virtual charging power of the station;

[0054] The total virtual charging power of the station is compared with the real-time available capacity of the station's main transformer. If the total virtual charging power continues to exceed the real-time available capacity for a preset time, it is determined that the transformer capacity limit constraint has been violated.

[0055] For each vehicle that is currently charging, predict the remaining charging time required to reach the target battery state of charge based on its current virtual charging power.

[0056] The predicted remaining charging end time is compared with the vehicle's scheduled charging end time. If the predicted end time is later than the end time, it is determined that the user's charging time constraint has been violated.

[0057] Check whether the virtual charging power of each charging station is between the minimum and maximum values ​​of its nameplate power;

[0058] Summarize all violations of constraints and impose appropriate penalties accordingly, or mark virtual joint actions as infeasible.

[0059] As a further aspect of the present invention, the generation of an orderly charging guidance scheme for a specific electric vehicle user includes:

[0060] The global optimal charging power allocation strategy is analyzed to obtain the total power budget allocated to the charging stations in each future time period and the power change trend;

[0061] Based on users' charging preference parameters, including expected vehicle pickup time, target battery capacity, and acceptable charging cost range, the user group is segmented.

[0062] For users who prefer low-cost charging and flexible charging times, a suggested off-peak charging plan is generated based on the time-of-use electricity price signal, and a charging start time suggestion can be pushed through the mobile application;

[0063] For users with urgent charging needs or who are time-sensitive, their charging power will be prioritized within the power budget range, and estimated full charge time information will be generated and pushed to the user.

[0064] Based on the ultra-short-term charging load forecast results, the power shortage period that the station is about to experience is dynamically assessed, and flexible guidance requests to wait or delay charging are sent to users who are about to start charging during the shortage period.

[0065] By combining the suggested charging start time, estimated full charge time information, and flexible guidance requests, a differentiated and orderly charging guidance scheme is formed for different user groups.

[0066] As a further aspect of the present invention, the segmentation of the user group includes:

[0067] Establish a user profile model with multiple dimensions, including historical charging behavior characteristics, sensitivity to electricity prices, time flexibility score, and vehicle battery capacity and current state of charge;

[0068] For each electric vehicle user who connects to the charging station, their scores in each dimension of the user profile model are calculated based on their current charging preference parameters and historical charging data.

[0069] Clustering algorithms, including K-means clustering, are applied to perform cluster analysis on all users currently in the station or who have made reservations, grouping users with similar profile characteristics into the same group;

[0070] Define a guidance strategy template for each user group, which includes the default power allocation priority, acceptable scheduling flexibility, and preferred information push method;

[0071] When generating a specific guidance plan, first determine the user's group, call the guidance strategy template corresponding to the group, and then combine it with the current global optimal charging power allocation strategy to generate personalized guidance content.

[0072] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0073] By deploying sensing units to systematically collect real-time status data of charging piles, battery status of connected vehicles, time-of-use electricity prices, and regional load data, a multi-dimensional operational dataset covering "source-grid-load-storage" is constructed. This enables the load forecasting model to integrate grid price signals, physical equipment status, and user behavior characteristics, changing the traditional method's reliance solely on historical total load curves. Based on this multi-dimensional data fusion, the generated short-term and ultra-short-term charging load forecasts have a more complete input feature space, enabling a more accurate depiction of the temporal changes and power distribution of charging demand at charging stations in future periods. This provides a near-real-world input basis for subsequent high-precision optimization decisions, reducing scheduling risks caused by forecast bias.

[0074] An improved multi-agent deep deterministic policy gradient algorithm is adopted as the decision-making core. Each agent can correspond to a charging unit or vehicle agent, and learns the optimal strategy through cooperation and competition. The algorithm innovatively encodes real-time time-of-use electricity price signals, transformer capacity limits, vehicle reservation time windows, and user preference parameters into a set of dynamically updated constraints, which are then integrated into the agent's objective function and state space. This design enables the optimization framework to directly perceive and respond to real-time changes in external electricity price fluctuations, internal equipment limits, and personalized needs. Through end-to-end training, the algorithm can directly and continuously calculate the future-period globally optimal power allocation strategy that satisfies all current dynamic constraints from a high-dimensional, continuous state-action space. Its decision-making process can adapt to environmental changes, avoiding the lag and complexity of traditional static models that require remodeling and solving when operating conditions change.

[0075] The globally optimal charging power allocation strategy generated through centralized optimization is transformed into individual power control commands for each charging station through a decomposition mechanism. This decomposition process is not a simple average allocation, but rather combines real-time battery status information of specific vehicles with user-preset charging preference parameters. Based on the vehicle's current battery level, battery health, target charging amount, and acceptable time range, the power commands are translated into specific charging time period suggestions, estimated completion time, and cost information, forming a personalized and orderly charging guidance plan. This establishes a transparent bridge between macro-optimization results and micro-user execution, enabling the optimization goals at the station level to be communicated and executed in a way that is understandable and acceptable to users, thereby improving user compliance and overall collaborative efficiency of the scheduling plan. Attached Figure Description

[0076] Figure 1 This is a state diagram of the load dispatching method for new energy power plants described in this invention;

[0077] Figure 2 A flowchart for the improved multi-agent deep deterministic policy gradient algorithm;

[0078] Figure 3 A flowchart for guiding network updates in the evaluation network of an agent. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0080] See Figure 1 This invention provides a load dispatching method for new energy power plants, the specific method including:

[0081] Sensing units deployed at new energy power stations collect multi-dimensional operational data, including real-time status of charging piles, battery status information of connected electric vehicles, time-of-use (TOU) pricing signals from the power grid, and real-time load data for the power station's supply area. Based on this multi-dimensional operational data, a load forecasting model is constructed and run to generate short-term and ultra-short-term charging load forecasts for the new energy power stations in future time periods. Using these forecasts as input, an improved multi-agent deep deterministic policy gradient algorithm is employed in the decision core to construct a scheduling optimization framework with the optimization objectives of reducing power station operating costs, smoothing power grid load fluctuations, and meeting user charging needs. The improved multi-agent deep deterministic policy gradient algorithm calculates a dynamic set of constraints based on real-time updated TOU pricing signals, the capacity limitations of the power station's main transformer, vehicle reservation charging times, and user-defined charging preference parameters. The scheduling optimization framework solves under this dynamic set of constraints, continuously outputting the globally optimal charging power allocation strategy for the next optimization cycle. The global optimal charging power allocation strategy is decomposed into individual power control commands for each charging station, and an orderly charging guidance scheme for specific electric vehicle users is generated based on vehicle battery status information and user charging preference parameters.

[0082] In one embodiment of the present invention, multi-dimensional operational data is preprocessed. The preprocessing includes removing outlier data points, filling in missing data values, and time-series alignment of data at different time scales. Charging load-related features are extracted from the preprocessed data. These features include historical load curves for the same period, real-time number of charging vehicles, average battery state of charge per vehicle, and spatiotemporal distribution information of future scheduled charging orders. These features are input into a short-term prediction network, which employs a convolutional neural network combined with gated recurrent units. Using an hourly time step, the network predicts the total charging load curve of new energy power stations for the next 24 hours, generating a short-term charging load prediction result. The features are then input into an ultra-short-term prediction network, which employs a temporal convolutional network structure. Using a minute-based time step, the network continuously predicts the total charging load curve of new energy power stations for the next 15 minutes to one hour, generating an ultra-short-term charging load prediction result. (See also...) Figure 2 Each charging pile within a new energy power station is defined as an agent, and an independent execution network and evaluation network are constructed for it. All agents share a global environmental state, which is composed of short-term charging load prediction results, ultra-short-term charging load prediction results, and time-of-use pricing signals. Each agent's execution network outputs the suggested charging power for its corresponding charging pile at the next time step based on the current global environmental state. The suggested charging power of all agents constitutes a joint action. The joint action is input into a centralized critic network, which evaluates the long-term expected total reward of the joint action based on the global environmental state and the joint action, and outputs a global state-action value function. Each agent's evaluation network uses the global state-action value function, combined with its local observation information, to calculate the merits of its actions, guiding the parameter updates of its execution network. At each training time step, the agent samples a batch of historical experience data from the experience replay pool. This historical experience data includes historical environmental states, historical joint actions, immediate rewards obtained, and new environmental states. Using the sampled historical experience data, the parameters of the centralized critic network are updated by minimizing the time difference error. The parameters of the execution network for each agent are updated via the policy gradient method along the direction of the global state-action value function evaluated by the commentator network in the boost set.

[0083] In practical implementation, a new energy power station equipped with 20 DC charging piles uses sensing units deployed at the station to collect multi-dimensional operational data. Real-time status data of the charging piles is collected once per minute, including charging power, voltage, current, and connection status. Battery status information of the connected electric vehicles is obtained through the vehicle-charging pile communication protocol, including battery state of charge, battery temperature, and health status. The time-of-use electricity price signal from the power grid is updated every 15 minutes by the power trading platform. Real-time load data of the power station's supply area is recorded every 15 minutes by the station's energy management system. During implementation, the raw operational data contains abnormal data points caused by brief communication interruptions. For example, a charging pile might have a power value of 0 at three consecutive sampling points, while the power values ​​in the preceding and following time periods are within the normal range. After preprocessing, these 0-value points are identified as abnormal and removed. For individual missing battery state of charge data due to network latency, linear interpolation of adjacent time points is used to fill the gaps. Data from different sources has different timestamps; the preprocessing process aligns all data to a minute-based time series. In practice, charging load-related features are extracted from time-series aligned data. Historical load curve features are used to extract the average and standard deviation of charging power at the same time over the past 7 days. Real-time charging vehicle quantity features are used to count the number of charging piles in charging and connection-ready states. Vehicle average battery state of charge features are used to calculate the average battery state of charge of all connected vehicles. The spatiotemporal distribution information of future scheduled charging orders is extracted from the reservation system, including the number of orders scheduled to arrive in the next 2 hours, the expected charging amount, and the reservation time period.

[0084] In some embodiments, the short-term prediction network employs a structure combining a convolutional neural network (CNN) and gated recurrent units (GRUs). The CNN contains two convolutional layers for extracting local features, and the GRU contains two recurrent units for learning long-term temporal dependencies. The network operates on an hourly time step, taking the feature sequence of the past 48 hours as input and outputting a predicted value for the total charging load curve of new energy power stations for the next 24 hours, thus generating a short-term charging load prediction result. In some embodiments, the ultra-short-term prediction network employs a temporal convolutional network structure, which includes dilated causal convolutional layers and residual connections. Operating on a minute-by-minute time step, it takes the feature sequence of the past 30 minutes as input and continuously predicts the total charging load curve of new energy power stations for the next 15 minutes to 1 hour, generating an ultra-short-term charging load prediction result. Data comparison is reflected in the analysis of the difference between the predicted and actual values. For example, on a certain operating day, the average absolute percentage error between the short-term charging load prediction result and the actual total charging load curve for that day is 8.5%, while the average absolute error of the ultra-short-term charging load prediction result when continuously predicting the next 30 minutes is 15 kilowatts.

[0085] In practical implementation, each charging pile within a new energy power station is defined as an intelligent agent, and an independent execution network and evaluation network are constructed for each agent. The execution network is a 3-layer fully connected neural network, with the number of input layer nodes consistent with the global environment state dimension, and the number of output layer nodes is 1, representing the suggested charging power value. The evaluation network is also a 3-layer fully connected neural network, with the number of input layer nodes being the sum of the global environment state dimension and the joint action dimension, and the number of output layer nodes is 1, representing the value assessment. All agents share a global environment state, which is a vector composed of short-term charging load prediction results, ultra-short-term charging load prediction results, and time-of-use electricity price signals. Specifically, the short-term charging load prediction results are processed into a vector of length 24, the ultra-short-term charging load prediction results are processed into a vector of length 12, and the time-of-use electricity price signal is processed into a vector of length 36 containing current and future electricity price information. The total dimension of the global environment state is 72. Each agent's execution network, based on the current global environment state vector, calculates a scalar value through forward propagation. This scalar value represents the suggested charging power of the corresponding charging station at the next moment. The outputs of the 20 agents collectively constitute a joint action vector with a dimension of 20. In practice, the joint action vector is concatenated with the global environment state vector and input into a centralized critic network. This centralized critic network is a 4-layer fully connected neural network. Based on the input global environment state and joint action, the network calculates through hidden layers and outputs a scalar. This scalar is the global state-action value function that evaluates the long-term expected total reward of the joint action. Each agent's evaluation network utilizes the global state-action value function output from the centralized critic network and combines it with the agent's local observation information, including the real-time occupancy status of the corresponding charging station, the battery state of the currently serving vehicle, the vehicle's charging duration, and the user's charging preference parameters. The evaluation network integrates this information to calculate the merits of the agent's actions, which guides the parameter updates of its execution network.

[0086] In practice, at each training time step, each agent samples a batch of historical experience data from a shared experience replay pool. The experience replay pool has a capacity of 10,000 experiences, and the batch size is 128. Each historical experience data item contains a historical environment state vector, a historical joint action vector, an acquired immediate reward scalar, and a new environment state vector. Using the sampled historical experience data, the parameters of the centralized critic network are updated by minimizing the temporal difference error. The calculation formula is:

[0087]

[0088] Where: symbol Indicates an immediate reward, symbol Represents the discount factor, symbol Indicates the current environmental state, symbol Indicates the current joint action, symbol Indicates the next environmental state, symbol The target action for the next state is indicated by the symbol. This represents the estimated value of the current state-action value function, with the symbol... This represents the target network's estimate of the value of the next state-action. The parameters of each agent's execution network are updated via the policy gradient method, along the direction of the global state-action value function evaluated by the commentator network in the boost set. The formula for calculating the policy gradient update is based on the deterministic policy gradient theorem.

[0089] Optionally, both the execution network and the evaluation network adopt the target network structure to improve training stability, and the parameters of the target network are periodically copied from the online network through soft updates. Optionally, exploration noise is added to the agent during the action selection phase. The exploration noise is generated using the Ornstein-Uhlenbeck process to enable effective exploration in the continuous action space.

[0090] In one embodiment of the present invention, see [reference] Figure 3The evaluation network receives the global state-action value function output from a centralized network of critics, using it as a global benchmark for evaluating the agent's actions. Simultaneously, the evaluation network receives local observation information from the agent, including the real-time occupancy status of the corresponding charging station, the battery state of the currently serving vehicle, the vehicle's charging duration, and the user's charging preference parameters. The global state-action value function and the agent's local observation information are fused and input into the multi-layer neural network of the evaluation network. Through forward propagation calculations in the multi-layer neural network, a local state-action value function estimate is output. This estimate represents the long-term value of executing the network's suggested action under the current global environment state and the agent's local observations. The difference between the state-action value function estimate and the sum of the actual immediate reward and the next state value estimate is calculated to obtain the temporal difference error for the agent. Using this temporal difference error, the network parameters of the evaluation network are updated via gradient backpropagation to reduce its error in evaluating action value. Based on the updated evaluation network, the value gradient of the actions suggested by the execution network is calculated. This value gradient indicates how actions should be adjusted to improve long-term returns. The value gradient is then passed to the execution network, which adjusts its network parameters using a policy gradient ascent method to make it more inclined to output actions that yield higher evaluations. The design of the immediate reward is determined based on the optimization objective and a set of dynamic constraints, including: calculating the total electricity cost of the new energy power station over a given time period after executing the joint actions, based on time-of-use pricing signals and total charging power, and generating a cost reward component negatively correlated with the total electricity cost; calculating the smoothness index of the total charging power curve of the new energy power station, which is the sum of squares of the rate of change of total charging power over the time period, and generating a smoothing reward component negatively correlated with the smoothness index; statistically analyzing the ratio of actual charging volume of all charging vehicles to the user's expected charging demand over the time period, and generating a demand satisfaction reward component positively correlated with the satisfaction ratio; and checking whether the joint actions violate the set of dynamic constraints, imposing a large negative penalty on actions that violate transformer capacity limits, and imposing a penalty that increases with time urgency on actions that violate user-set charging deadline constraints. The cost reward component, smoothing reward component, demand satisfaction reward component, and various penalties are weighted and summed. The result is used as the instantaneous reward obtained by all agents at each time step, and is fed back to the improved multi-agent deep deterministic policy gradient algorithm.

[0091] In practice, the evaluation network receives the global state-action value function output from a centralized network of critics. This global state-action value function is a scalar value, which the evaluation network uses as a global benchmark reference for evaluating the agent's actions. Simultaneously, the evaluation network also receives local observation information from the agent. For an agent connected to a charging station for an electric vehicle, this local observation information includes the charging station's real-time occupancy status ("charging"), the current vehicle's battery charge level (65%), the vehicle's charging duration (25 minutes), and the user's charging preference parameters, including a desired charge to 90% and a "high" charging cost sensitivity level. The evaluation network fuses the global state-action value function with the agent's local observation information. The global state-action value function scalar is copied and expanded into a multi-dimensional feature vector, which is then concatenated with the agent's local observation information vector to form a fused feature vector, which is input into the multi-layer neural network of the evaluation network. The evaluation network performs forward propagation computation using a three-layer fully connected neural network, outputting a local state-action value function estimate. This estimate represents the long-term value of executing the network's suggested action given the current global environment state and the agent's local observations. In practice, the difference between the local state-action value function estimate and the sum of the actual immediate reward and the next state value estimate (calculated by the target evaluation network) is calculated. This difference represents the temporal difference error specific to the agent.

[0092] Using temporal difference errors, the network parameters of the evaluation network are updated via gradient backpropagation. Weight updates follow the gradient descent principle to reduce the evaluation network's error in assessing action value. Based on the updated evaluation network, the value gradient of the actions suggested by the execution network is calculated. The value gradient is the partial derivative of the evaluation network's output with respect to the execution network's output action, indicating how actions should be adjusted to improve long-term rewards. The value gradient is then passed to the execution network, which adjusts its network parameters using a policy gradient ascent method, making it more inclined to output actions that yield higher evaluations. The parameter update formula for the execution network in a training batch is:

[0093]

[0094] Where: symbol The objective function representing the policy performance with respect to the execution network parameters. Gradient estimation, sign Indicates the number of samples collected, symbol This represents the summation operation on the samples, with the symbol... This represents the local state-action value function estimate of the network output with respect to the execution of the network output action. gradient, sign Indicates the execution of network policies Regarding its own parameters gradient, sign This represents the agent's observation. In a simulated scenario, when the execution network suggests a charging power of 120 kW, and the evaluation network calculates a higher value based on global and local information, the value gradient is positive. The execution network will then fine-tune its internal parameters to make it more likely to output a power value close to or higher than 120 kW in similar situations in the future.

[0095] In some embodiments, the design of the instant reward is determined based on the optimization objective and the set of dynamic constraints. After the joint action is executed, the total electricity cost of the new energy power station in a time period is calculated based on the time-of-use electricity price signal and the total charging power. In a calculation cycle, the time-of-use electricity price signal shows an electricity price of 0.8 yuan / kWh, the total charging power of the power station is 1500 kW, and the time period is 0.25 hours. Then the total electricity cost is 0.8 * 1500 * 0.25 = 300 yuan. The cost reward component is set as the negative value of the total electricity cost divided by a scaling factor, for example, -300 / 100 = -3. In some embodiments, the smoothness index of the total charging power curve of the new energy power station is calculated. The smoothness index is the sum of squares of the rate of change of the total charging power within a time period. In an optimization cycle containing 4 time steps, the total charging power sequence is [1500, 1520, 1480, 1510] kW, the time interval is 15 minutes, the rate of change sequence is [20, -40, 30] kW / 15 minutes, the smoothness index is 20² + (-40)² + 30² = 2900, and the smoothing reward component is set as the negative value of the smoothness index divided by a scaling factor, for example, -2900 / 1000 = -2.9. The statistics show the ratio of the actual charging amount of all charging vehicles to the user's expected charging demand within a given time period. Assuming there are 10 vehicles charging, with a total actual charging amount of 300 kWh, and the user's total expected charging demand is 320 kWh, the ratio is 300 / 320 = 0.9375. The reward for meeting the demand is set as the ratio multiplied by a scaling factor, for example, 0.9375 * 10 = 9.375.

[0096] Optionally, check whether the combined actions violate the set of dynamic constraints. Apply a large negative penalty to actions that violate transformer capacity limits. For example, if the main transformer capacity limit at the station is 2000 kW, and the combined actions result in a total charging power of 2100 kW, apply a penalty of -50. Alternatively, apply a penalty that increases with time urgency to actions that violate user-set charging deadline constraints. For a vehicle required to leave in 60 minutes, if its charging completion is predicted to take 80 minutes, the time urgency is 20 minutes, and the penalty value can be set to -1.5 * 20 = -30. The cost reward component, smoothing reward component, demand satisfaction reward component, and various penalties are weighted and summed. The weight coefficients can be set according to the operation strategy. For example, the cost weight is 0.5, the smoothing weight is 0.2, and the demand satisfaction weight is 0.3. The result is used as the instant reward obtained by all agents at the time step and fed back to the improved multi-agent deep deterministic policy gradient algorithm. In the above example, if there is no constraint violation, the instant reward is calculated as 0.5*(-3)+0.2*(-2.9)+0.3*9.375=1.0325.

[0097] It is understandable that the evaluation network's computation relies on a global value signal provided by a centralized network of critics, ensuring that individual agents' policy updates consider the overall system reward. It is also understandable that the weighting of each component in the immediate reward function directly determines the balancing tendency of the scheduling optimization framework among multiple objectives such as reducing power station operating costs, smoothing grid load fluctuations, and meeting user charging needs.

[0098] In one embodiment of the present invention, time-of-use electricity price signals are acquired in real time. The electricity price change time points within the future optimization cycle are determined based on the electricity price time period division boundaries. Power cost weight vectors for different time periods are generated based on these price change time points. The current load rate and rated capacity of the main transformer at the power station are acquired in real time. The real-time available capacity of the transformer at each moment within the optimization cycle is calculated, and the changing trend of capacity limits is predicted by combining historical load curves, generating a time-indexed sequence of capacity limits. All connected and reserved vehicles are traversed, and the reserved charging deadline time and current time for each vehicle are extracted. The remaining available time window for each vehicle's charging demand is calculated. The minimum average charging power requirement for each vehicle is generated by combining the vehicle's current battery state of charge and the target battery state of charge. The user's set charging preference parameters are analyzed, and the user's acceptable latest completion time, expected power increase range, and sensitivity level to charging costs are extracted. The latest completion time is compared with the reserved charging deadline time, and the more stringent time is taken as the user's effective charging deadline constraint. By integrating the power cost weight vector, capacity limit sequence, minimum average charging power requirement of each vehicle, and effective charging cutoff time constraint of each user, a dynamic constraint mathematical expression containing inequality constraints and equality constraints is constructed. The dynamic constraint set is composed of the real-time calculation results of the dynamic constraint mathematical expression.

[0099] In practical implementation, the improved multi-agent deep deterministic strategy gradient algorithm calculates a dynamic set of constraints based on real-time updated time-of-use (TOU) electricity price signals, the capacity limitations of the main transformer at the power station, the scheduled charging time of vehicles, and user-defined charging preference parameters. Specifically, the TOU electricity price signals released by the power grid are acquired in real time, updated at 15-minute intervals. For example, the signal acquired at 3 PM on a certain workday shows a flat price of 0.8 yuan / kWh from 3 PM to 5 PM, a peak price of 1.2 yuan / kWh from 5 PM to 8 PM, and a valley price of 0.5 yuan / kWh after 8 PM. The electricity price change time points within the future optimization cycle are determined based on the electricity price time period boundaries. The optimization cycle is set to the next 4 hours, with the corresponding price change time points being 5 PM and 8 PM. Power cost weight vectors for different time periods are generated based on these price change time points. These power cost weight vectors are proportional to the electricity price and are used to reflect electricity costs in the optimization objective. An example power cost weight vector is shown in Table 1.

[0100] Table 1: Power Cost Weighting Table by Time Period During the Optimization Period

[0101]

[0102] In practice, the current load rate and rated capacity of the main transformer at the power station are acquired in real time. The rated capacity of the main transformer is 2500kVA, and the current load power is 800kW. The real-time available capacity of the transformer at the current moment is calculated to be 2500-800=1700kW. Combining historical load curves to predict the changing trend of capacity limits, historical data shows that the output of the power station's own photovoltaic power generation decreases in the evening, leading to an increase in the net load that the transformer needs to bear. Therefore, a capacity limit sequence is generated with 15-minute intervals for the next 4 hours. The value of the sequence gradually decreases from 1700kW to 1500kW between 5 pm and 7 pm to reflect the tightening trend of available capacity.

[0103] In some embodiments, all connected and reserved vehicles are iterated. Assume there are currently 8 electric vehicles charging, and another 5 have been reserved to arrive within the next 2 hours. The reserved charging deadline and current time for each vehicle are extracted. For a vehicle currently charging, its reserved charging deadline is 5:30 PM, and the current time is 3:00 PM, so the remaining available time window is 150 minutes. Combining the vehicle's current battery state of charge (SOC) and target SOC, the current SOC is 40%, the user sets the target SOC to 90%, and the battery capacity is 60 kWh. Therefore, the required charging amount is (90% - 40%) * 60 = 30 kWh. The formula for generating the minimum average charging power requirement for each vehicle is:

[0104]

[0105] Where: symbol Indicates the minimum average charging power required by the vehicle, symbol Indicates the remaining amount of charge required by the vehicle, symbol This represents the remaining available time window from the current moment until the charging deadline. For the example vehicle above, the minimum average charging power requirement is 30kWh / 2.5h = 12kW. Perform this calculation for all vehicles to obtain a list of minimum average charging power requirements for each vehicle.

[0106] Optionally, the system parses user-defined charging preference parameters. These parameters, input by the user via the mobile application, include "Last completion time: 6 PM," "Desired battery level: 90%," and "Cost sensitivity: High." The system extracts the user's acceptable latest completion time, desired battery level increase range, and cost sensitivity level. It then compares the latest completion time with the scheduled charging deadline obtained from the order system, selecting the stricter time as the user's effective charging deadline constraint. For example, if the scheduled charging deadline is 5:30 PM, and the user's latest completion time is 6 PM, then the effective charging deadline constraint is 5:30 PM. The cost sensitivity level will be used to configure different cost penalty weights in the optimization objective.

[0107] It is understandable that the power cost weight vector directly affects the scheduling algorithm's trade-off between cost and other objectives; a high weight during peak periods will guide the algorithm to reduce power during those periods if possible. The minimum average charging power demand is a hard constraint, ensuring that the minimum required charging amount can be completed within the user-defined time window. By integrating the power cost weight vector, capacity limit sequence, minimum average charging power demand for each vehicle, and effective charging deadline constraints for each user, a dynamic constraint mathematical expression containing inequality and equality constraints is constructed. The dynamic constraint set consists of the real-time calculation results of these mathematical expressions. For example, a basic constraint expression includes: the sum of the power of all charging piles must not exceed the transformer's capacity limit at any given time; the actual average charging power of each vehicle must not be lower than its calculated minimum average charging power demand; and the charging completion time of each vehicle must not be later than its effective charging deadline constraint.

[0108] In one embodiment of the present invention, at the beginning of each optimization cycle, the states of all agents in the improved multi-agent deep deterministic policy gradient algorithm are initialized. These states include the real-time states of each charging pile, the battery status information of each vehicle, the current electricity price period, and the predicted short-term charging load. The improved multi-agent deep deterministic policy gradient algorithm aims to maximize long-term cumulative rewards by exploring in a continuous action space and simulating charging power allocation scenarios for multiple future time steps. During the simulation, each time step is advanced, the current virtual joint action is substituted into a set of dynamic constraints for feasibility verification. If the verification passes, the virtual reward is calculated according to the reward function, the agent state is updated, and the simulation continues. If the verification fails, a constraint penalty is imposed on the joint action, guiding the agent to avoid such actions in subsequent simulations, and triggering local replanning. After completing a preset number of simulation exploration steps, the action sequence that obtains the highest cumulative reward is selected as the optimization result of the current optimization cycle, and the joint action of the first time step of the action sequence is output as the initial part of the globally optimal charging power allocation strategy. The feasibility verification process is as follows: Extract the virtual joint action at the current time step, i.e., the suggested virtual charging power values ​​of all charging piles, and sum them to obtain the total virtual charging power of the charging station. Compare the total virtual charging power of the charging station with the real-time available capacity of the station's main transformer. If the total virtual charging power continuously exceeds the real-time available capacity for a preset time, it is determined that the transformer capacity limit constraint has been violated. For each vehicle currently charging, predict the remaining charging time required to reach the target battery state of charge based on its current virtual charging power. Compare the predicted remaining charging end time with the vehicle's scheduled charging deadline. If the predicted end time is later than the deadline, it is determined that the user charging time constraint has been violated. Check whether the virtual charging power of each charging pile is between the minimum and maximum values ​​of its nameplate power. Summarize all constraint violations and apply corresponding penalties, or mark the virtual joint action as infeasible.

[0109] In practical implementation, the scheduling optimization framework solves the problem under a set of dynamic constraints. At the beginning of each optimization cycle, the states of all agents in the improved multi-agent deep deterministic policy gradient algorithm are initialized. State information includes the real-time state of each charging pile (e.g., 5 charging piles are "charging" and 3 are "idle"), the battery state information of each vehicle (including battery state of charge and battery health), the current electricity price period ("flat"), and the predicted short-term charging load. In practice, the improved multi-agent deep deterministic policy gradient algorithm aims to maximize long-term cumulative rewards by exploring the continuous action space and simulating charging power allocation scenarios for multiple future time steps. The optimization cycle is set to the next 2 hours, with each time step lasting 15 minutes, for a total of 8 time steps. During the simulation, each time step is advanced, the current virtual joint action needs to be substituted into the set of dynamic constraints for feasibility verification. The virtual joint action is a vector containing the suggested charging power for all charging piles in the current simulation time step.

[0110] In some embodiments, if the verification passes, a virtual reward is calculated based on the reward function, and the agent's state is updated to continue the simulation. State updates include updating the remaining required charging amount for each vehicle, the occupancy status of charging stations, and the current electricity price period. In some embodiments, if the verification fails, a constraint penalty is imposed on the joint action, guiding the agent to avoid such actions in subsequent simulations. Local replanning may also be triggered, whereby the action exploration space of the agent triggering the constraint is adjusted in subsequent simulations, for example, limiting its output power to a smaller feasible range. After completing a preset number of simulation steps, the action sequence that yields the highest cumulative reward is selected as the optimization result for the current optimization cycle. The joint action of the first time step of the action sequence is output as the initial part of the globally optimal charging power allocation strategy. This initial part is a power command vector used to directly control the power of each charging station in the next 15-minute period.

[0111] In practical implementation, the feasibility verification process involves extracting the virtual joint action of the current time step, i.e., the virtual charging power suggestion values ​​of all charging piles, for example, a vector containing the power values ​​of 8 charging piles [120,0,90,150,0,60,120,0] (unit: kilowatts). The total virtual charging power of the station is summed; in the example above, the total power is 540 kilowatts. The total virtual charging power of the station is compared with the real-time available capacity of the station's main transformer. The real-time available capacity of the station's main transformer at the current simulation time is 600 kilowatts, and the total virtual charging power of 540 kilowatts does not exceed this limit. If the total virtual charging power continuously exceeds the real-time available capacity for a preset time, it is determined that the transformer capacity constraint has been violated. The preset time can be set to two consecutive simulation time steps. In another scenario, the simulation sequence shows that the total virtual charging power of two consecutive time steps is 610 kilowatts and 620 kilowatts, respectively, both exceeding the real-time available capacity of 600 kilowatts, thus triggering a constraint violation. See Table 2 for a transformer capacity constraint check.

[0112] Table 2: Transformer Capacity Constraint Checklist During Simulation

[0113]

[0114] Optionally, for each vehicle currently charging, based on its current virtual charging power, predict the remaining charging time required to reach the target battery state of charge. The calculation formula is:

[0115]

[0116] Where: symbol This indicates the predicted remaining charging time, in hours, with the symbol [symbol missing]. Indicates the target battery state of charge of the vehicle, symbol This indicates the battery state of charge of the vehicle at the current simulation moment, denoted by [symbol]. This indicates the vehicle's battery capacity, measured in kilowatt-hours (kWh), with the symbol... This represents the virtual charging power allocated to the corresponding charging station for the vehicle at the current time step, in kilowatts (kW). The predicted remaining charging end time is compared with the vehicle's scheduled charging deadline. If the predicted end time is later than the deadline, it is considered a violation of the user's charging time constraint. For example, if a vehicle's battery state of charge is 50% at the current simulation moment, the target is 80%, the battery capacity is 60 kWh, and the allocated power is 12 kW, the calculated... The time limit is 1.5 hours. If the reservation deadline is more than 1 hour later, the restriction will be considered a violation.

[0117] Optionally, check if the virtual charging power of each charging pile is between the minimum and maximum values ​​of its nameplate power. The nameplate power range for each charging pile is 0 to 180 kW. If the virtual charging power of a charging pile is 200 kW, it is determined to violate the physical constraints of the equipment. Summarize all constraint violations and apply corresponding penalties, or mark virtual joint actions as infeasible. During the simulation, if a joint action is marked as infeasible, the state transition corresponding to that action will not be added to the experience replay pool for subsequent training. It is understood that feasibility verification is a crucial step in the simulation exploration, ensuring that the action space explored by the agent conforms to physical and business constraints. It is also understood that penalizing actions that violate constraints and guiding subsequent exploration is an effective way to integrate hard constraints into the reinforcement learning training framework, enabling the algorithm to autonomously learn to find the optimal strategy within the constraint boundaries.

[0118] In one embodiment of the invention, a globally optimal charging power allocation strategy is analyzed to obtain the total power budget allocated to the charging stations in future time periods and the power change trend. Based on user charging preference parameters, including expected vehicle pickup time, target battery capacity, and acceptable charging cost range, user groups are segmented. For users who prefer low-cost charging and flexible charging times, a suggested off-peak charging plan is generated based on time-of-use pricing signals, and charging start time suggestions can be pushed via a mobile application. For users with urgent charging needs or time sensitivity, their charging power is prioritized within the power budget, and estimated full charge time information is generated and pushed to them. Based on ultra-short-term charging load forecasts, upcoming power shortage periods at the charging stations are dynamically assessed, and flexible guidance requests for waiting or delayed charging are sent to users who are about to start charging during these periods. By integrating the suggested charging start time, estimated full charge time information, and flexible guidance requests, a differentiated and orderly charging guidance plan is formed for different user groups. A user profile model with multiple dimensions is established, including historical charging behavior characteristics, sensitivity to electricity prices, time flexibility score, and vehicle battery capacity and current state of charge. For each electric vehicle user accessing the charging station, their score across various dimensions of the user profile model is calculated based on their current charging preference parameters and historical charging data. Clustering algorithms, including K-means clustering, are applied to analyze all users currently at the charging station or those who have made reservations, grouping users with similar profile characteristics into the same group. A guidance strategy template is defined for each user group, including a default power allocation priority, acceptable scheduling flexibility, and preferred information push methods. When generating a specific guidance plan, the user's group is first determined, the corresponding guidance strategy template is invoked, and then combined with the current globally optimal charging power allocation strategy to generate personalized guidance content.

[0119] In practical implementation, an orderly charging guidance scheme is generated for specific electric vehicle users. The globally optimal charging power allocation strategy is analyzed to obtain the total power budget allocated to the charging stations in each future time period and the power change trend. The globally optimal charging power allocation strategy outputs the power plan for the next 2 hours at 15-minute intervals. For example, a power plan sequence is [500, 520, 480, 510, 530, 550, 570, 590] (unit: kilowatts). Based on the user's charging preference parameters, including expected vehicle pickup time, target battery level, and acceptable charging cost range, the user group is segmented. The user's charging preference parameters are submitted through a mobile application. For example, user A sets the expected vehicle pickup time to 2 hours later, the target battery level to 90%, and the cost range to "low cost priority," while user B sets the expected vehicle pickup time to 30 minutes later, the target battery level to 80%, and the cost range to "time priority."

[0120] In practice, for users who prefer low-cost charging and have flexible charging times, the system generates suggested off-peak charging plans based on time-of-use (TOU) pricing signals. The TOU signal indicates a current peak price of 1.2 yuan / kWh, with the off-peak price at 0.5 yuan / kWh expected after one hour. The system then pushes a charging start time suggestion to identified users with flexible charging times, stating, "We recommend starting charging in one hour; this is expected to save 65% on electricity costs." For users with urgent charging needs or who are time-sensitive, their charging power is prioritized within the power budget, and an estimated full charge time is generated and pushed to them. For user B, based on their vehicle's current status and the allocated higher power, the system calculates an estimated full charge time of 25 minutes and pushes a message stating, "You will be prioritized; charging is expected to be completed in 25 minutes." Based on the ultra-short-term charging load forecast, the system dynamically assesses the upcoming power shortage periods at charging stations. The forecast indicates that total charging demand will exceed 90% of the current power budget in the next 30 minutes. The system sends a flexible guidance request to user C, who has already scheduled to start charging in 25 minutes. The request states, "The charging station is about to become busy. Would you be willing to delay your charging start by 15 minutes? We will reserve a charging spot for you and provide a points reward." By combining the suggested charging start time, estimated charging time, and the flexible guidance request, a differentiated and orderly charging guidance plan is formed for different user groups. This differentiation is reflected in pushing definitive time information to time-sensitive users, offering cost-saving suggestions to price-sensitive users, and providing flexible adjustment requests to users with greater flexibility.

[0121] In some embodiments, a user profile model with multiple dimensions is established, including historical charging behavior characteristics, sensitivity to electricity prices, time flexibility score, and vehicle battery capacity and current state of charge. Historical charging behavior characteristics are extracted from the user's past charging records, including average charging duration, frequently used charging periods, and average charging amount. Sensitivity to electricity prices is quantified by analyzing the user's historical charging ratio during peak and off-peak hours. The time flexibility score is calculated based on the frequency and advance notice of the user's historical modifications or cancellations of reservation orders. Vehicle battery capacity and current state of charge are real-time vehicle information. In some embodiments, for each electric vehicle user accessing charging, based on their current charging preference parameters and historical charging data, their score on each dimension of the user profile model is calculated. For example, user D's historical charging behavior characteristic score is 0.8, their sensitivity to electricity prices score is 0.9, their time flexibility score is 0.7, and their current vehicle battery state of charge is 50%. These data constitute a feature vector.

[0122] Clustering algorithms, including K-means clustering, are applied to analyze all users currently at the facility or who have made reservations, grouping users with similar profile characteristics into the same group. During clustering, the weighted Euclidean distance between user feature vectors is calculated, using the following formula:

[0123]

[0124] Where: symbol Represents the feature distance between user x and user y, with the symbol... These represent the weights of four dimensions: historical charging behavior, electricity price sensitivity, time flexibility, and battery state, respectively. and Let x and y represent the standardized scores of users x and y respectively on the dimension of historical charging behavior, with symbols... and The electricity price sensitivity score is represented by the symbol. and The time flexibility score is indicated by the symbol. and This represents a standardized value indicating the battery's state of charge. Through iterative calculations, users are ultimately categorized into three groups: "highly price-sensitive and flexible," "highly time-sensitive and urgent," and "balanced."

[0125] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A load dispatching method for new energy power plants, characterized in that, include: The sensing units deployed at new energy power stations collect multi-dimensional operational data, including the real-time status of charging piles, battery status information of connected electric vehicles, time-of-use electricity price signals of the power grid, and real-time load data of the power station's power supply area. Based on the multi-dimensional operational data, a load forecasting model is constructed and run to generate short-term and ultra-short-term charging load forecasts for new energy power stations in future periods. Based on the short-term and ultra-short-term charging load forecast results, an improved multi-agent deep deterministic strategy gradient algorithm is used in the decision core to construct a scheduling optimization framework with the optimization objectives of reducing station operating costs, smoothing grid load fluctuations, and meeting user charging needs. The improved multi-agent deep deterministic strategy gradient algorithm calculates a dynamic set of constraints based on the real-time updated time-of-use electricity price signal, the capacity limit of the main transformer at the power station, the vehicle's scheduled charging time, and the user's set charging preference parameters. The scheduling optimization framework solves the problem under the dynamic set of constraints and outputs the globally optimal charging power allocation strategy for the next optimization cycle. The global optimal charging power allocation strategy is decomposed into individual power control commands for each charging pile, and an orderly charging guidance scheme for specific electric vehicle users is generated based on vehicle battery status information and user charging preference parameters.

2. The load dispatching method for new energy power plants according to claim 1, characterized in that, Based on the aforementioned multi-dimensional operational data, a load forecasting model is constructed and run, including: The multi-dimensional operational data is preprocessed, including removing outlier data points, filling in missing data values, and aligning data at different time scales to time series. Charging load-related features are extracted from the preprocessed data. These features include historical load curves for the same period, the number of vehicles charging in real time, the average battery state of charge of the vehicles, and the spatiotemporal distribution information of future scheduled charging orders. The features are input into a short-term prediction network, which uses a convolutional neural network combined with a gated recurrent unit structure. The network uses hours as the time step to predict the total charging load curve of new energy power stations for the next 24 hours and generates the short-term charging load prediction result. The features are input into an ultra-short-term prediction network, which adopts a temporal convolutional network structure and uses minutes as the time step to predict the total charging load curve of new energy power stations for the next fifteen minutes to one hour, thereby generating the ultra-short-term charging load prediction result.

3. The load dispatching method for new energy power plants according to claim 1, characterized in that, The improved multi-agent deep deterministic policy gradient algorithm has the following working steps: Each charging pile in a new energy power station is defined as an intelligent agent, and an independent execution network and evaluation network are built for it. All intelligent agents share a global environmental state, which is composed of the short-term charging load prediction results, the ultra-short-term charging load prediction results, and the time-of-use electricity price signal. Each agent's execution network outputs the suggested charging power of its corresponding charging station in the next moment based on the current global environment state. The suggested charging power of all agents together constitutes a joint action. The joint action is input into a centralized network of critics, which evaluates the long-term expected total reward of the joint action based on the global environment state and the joint action, and outputs a global state-action value function. Each agent's evaluation network uses the global state-action value function and combines it with the agent's local observation information to calculate the quality of the agent's actions, which is used to guide the parameter updates of its execution network. At each training time step, the agent samples a batch of historical experience data from the experience replay pool, which includes historical environmental states, historical joint actions, immediate rewards obtained, and new environmental states. Using the sampled historical experience data, the parameters of the network of critics in the set are updated by minimizing the temporal difference error; The parameters of the execution network for each agent are updated via a policy gradient method in the direction that improves the global state-action value function evaluated by the commentator network in the set.

4. The load dispatching method for new energy power plants according to claim 3, characterized in that, The evaluation network for each agent uses the global state-action value function, combined with the agent's local observation information, to calculate the quality of the agent's actions, which guides the parameter updates of its execution network, including: The evaluation network receives the global state-action value function output from the network of critics in the set, and uses it as a global benchmark reference value for evaluating the actions of the agent. Meanwhile, the evaluation network receives local observation information of the agent, which includes the real-time occupancy status of the corresponding charging pile, the battery charge status of the currently served vehicle, the charging time of the vehicle, and the user's charging preference parameters. The global state-action value function is fused with the local observation information of the agent, and then input into the multi-layer neural network of the evaluation network. The evaluation network is computed through forward propagation of a multi-layer neural network to output a local state-action value function estimate. The state-action value function estimate represents the long-term value of executing the actions suggested by the network under the current global environment state and the local observations of the agent. The difference between the estimated state-action value function and the actual immediate reward plus the estimated value of the next state is calculated to obtain the temporal difference error for this agent. The network parameters of the evaluation network are updated using the temporal difference error through the gradient backpropagation algorithm to reduce its error in action value assessment. Based on the updated evaluation network, the value gradient of the action suggested by the execution network is calculated, and the value gradient indicates how the action should be adjusted to improve long-term returns; The value gradient is passed to the execution network, which then uses a policy gradient ascent method to adjust its network parameters based on this gradient, making the execution network more inclined to output actions that can obtain higher evaluations.

5. The load dispatching method for new energy power plants according to claim 3, characterized in that, The design of the immediate reward is determined based on the optimization objective and the set of dynamic constraints, including: After the joint action is executed, the total electricity cost of the new energy power station in a certain period of time is calculated based on the time-of-use electricity price signal and the total charging power, and a cost reward component that is negatively correlated with the total electricity cost is generated. Calculate the smoothness index of the total charging power curve of new energy power stations. The smoothness index is the sum of squares of the rate of change of total charging power within a time period, and generate a smoothing reward component that is negatively correlated with the smoothness index. The system calculates the ratio of the actual charging amount of all charging vehicles to the user's expected charging demand within a given time period, and generates a demand fulfillment reward component that is positively correlated with the fulfillment ratio. Check whether the combined action violates the set of dynamic constraints. Apply a large negative penalty to actions that violate the transformer capacity limit, and apply a penalty that increases with time urgency to actions that violate the user-set charging cutoff time constraint. The cost reward component, smoothing reward component, demand satisfaction reward component, and various penalties are weighted and summed. The result is used as the instantaneous reward obtained by all agents at each time step and fed back to the improved multi-agent deep deterministic policy gradient algorithm.

6. The load dispatching method for new energy power plants according to claim 1, characterized in that, The improved multi-agent deep deterministic policy gradient algorithm calculates a dynamic set of constraints based on the real-time updated time-of-use electricity price signal, the capacity limit of the main transformer at the power station, the vehicle's scheduled charging time, and the user-defined charging preference parameters. These constraints include: The time-of-use electricity price signal is acquired in real time, and the electricity price change time points in the future optimization cycle are determined according to the electricity price time period division boundary. Based on the electricity price change time points, power cost weight vectors for different time periods are generated. The current load rate and rated capacity of the main transformer in the power station are obtained in real time. The real-time available capacity of the transformer at each moment in the optimization cycle is calculated. The changing trend of capacity limit is predicted by combining the historical load curve and a time-indexed capacity limit sequence is generated. Iterate through all connected and reserved vehicles, extract the reserved charging deadline and current time for each vehicle, calculate the remaining available time window for each vehicle's charging needs, and combine the vehicle's current battery state of charge and target battery state of charge to generate the minimum average charging power requirement for each vehicle. The system analyzes the user's charging preference parameters, extracts the user's acceptable latest completion time, expected range of battery capacity increase, and sensitivity level to charging costs, compares the latest completion time with the scheduled charging deadline, and takes the stricter time as the user's effective charging deadline constraint. By integrating the power cost weight vector, the capacity upper limit sequence, the minimum average charging power requirement of each vehicle, and the effective charging cutoff time constraint of each user, a dynamic constraint mathematical expression containing inequality constraints and equality constraints is constructed. The dynamic constraint set is composed of the real-time calculation results of the dynamic constraint mathematical expression.

7. The load dispatching method for new energy power plants according to claim 1, characterized in that, The scheduling optimization framework solves the problem under the dynamic set of constraints, including: At the beginning of each optimization cycle, the states of all agents in the improved multi-agent deep deterministic policy gradient algorithm are initialized. The states include the current real-time state of each charging pile, the battery status information of each vehicle, the current electricity price period, and the future short-term charging load prediction results. The improved multi-agent deep deterministic policy gradient algorithm aims to maximize long-term cumulative rewards by exploring in the continuous action space and simulating charging power allocation scenarios for multiple future time steps. During the simulation, at each time step forward, the current virtual joint action needs to be substituted into the dynamic set of constraints for feasibility verification. If the verification passes, the virtual reward is calculated according to the reward function, the agent state is updated, and the simulation continues. If the verification fails, a constraint penalty is imposed on the joint action, and the agent is guided to avoid such actions in subsequent simulations. At the same time, local replanning may be triggered. After completing a preset number of simulated exploration steps, the action sequence that obtains the highest cumulative reward is selected as the optimization result of the current optimization cycle, and the joint action of the first time step of the action sequence is output as the initial part of the global optimal charging power allocation strategy.

8. The load dispatching method for new energy power plants according to claim 7, characterized in that, The feasibility verification process is specifically as follows: Extract the virtual joint action of the current time step, that is, the virtual charging power suggestion value of all charging piles, and sum them to obtain the total virtual charging power of the station; The total virtual charging power of the station is compared with the real-time available capacity of the station's main transformer. If the total virtual charging power continues to exceed the real-time available capacity for a preset time, it is determined that the transformer capacity limit constraint has been violated. For each vehicle that is currently charging, predict the remaining charging time required to reach the target battery state of charge based on its current virtual charging power. The predicted remaining charging end time is compared with the vehicle's scheduled charging end time. If the predicted end time is later than the end time, it is determined that the user's charging time constraint has been violated. Check whether the virtual charging power of each charging station is between the minimum and maximum values ​​of its nameplate power; Summarize all violations of constraints and impose appropriate penalties accordingly, or mark virtual joint actions as infeasible.

9. The load dispatching method for new energy power plants according to claim 1, characterized in that, The generation of an orderly charging guidance scheme for specific electric vehicle users includes: The global optimal charging power allocation strategy is analyzed to obtain the total power budget allocated to the charging stations in each future time period and the power change trend; Based on users' charging preference parameters, including expected vehicle pickup time, target battery capacity, and acceptable charging cost range, the user group is segmented. For users who prefer low-cost charging and flexible charging times, a suggested off-peak charging plan is generated based on the time-of-use electricity price signal, and a charging start time suggestion can be pushed through the mobile application; For users with urgent charging needs or who are time-sensitive, their charging power will be prioritized within the power budget range, and estimated full charge time information will be generated and pushed to the user. Based on the ultra-short-term charging load forecast results, the power shortage period that the station is about to experience is dynamically assessed, and flexible guidance requests to wait or delay charging are sent to users who are about to start charging during the shortage period. By combining the suggested charging start time, estimated full charge time information, and flexible guidance requests, a differentiated and orderly charging guidance scheme is formed for different user groups.

10. The load dispatching method for new energy power plants according to claim 9, characterized in that, The segmentation of user groups includes: Establish a user profile model with multiple dimensions, including historical charging behavior characteristics, sensitivity to electricity prices, time flexibility score, and vehicle battery capacity and current state of charge; For each electric vehicle user who connects to the charging station, their scores in each dimension of the user profile model are calculated based on their current charging preference parameters and historical charging data. Clustering algorithms, including K-means clustering, are applied to perform cluster analysis on all users currently in the station or who have made reservations, grouping users with similar profile characteristics into the same group; Define a guidance strategy template for each user group, which includes the default power allocation priority, acceptable scheduling flexibility, and preferred information push method; When generating a specific guidance plan, first determine the user's group, call the guidance strategy template corresponding to the group, and then combine it with the current global optimal charging power allocation strategy to generate personalized guidance content.

Images