Multi-zone-area charging scheduling method and system considering charging load change trend

Abstract

The invention relates to a multi-zone-area charging scheduling method and system considering a charging load change trend, and the method comprises the steps: carrying out the preprocessing of the historical charging load data of zone areas, and taking the data of a set time period as a sample; training a neural network model according to the obtained sample, and obtaining a load prediction curve in combination with the currently collected load data; according to the obtained load prediction curve, state information of load transfer of each transformer area, transformer area load information, change rate per unit time, voltage information and tie line load, load balancing, network loss minimization and charging pile utilization rate maximization are taken as targets, active power imbalance, reactive power imbalance and line overload are taken as constraints, and the load of each transformer area is predicted. And a multi-zone-area optimization strategy is obtained by using deep reinforcement learning, and load balancing is realized. Optimized scheduling is realized according to dynamic characteristics of loads, collaborative optimization of multiple transformer areas is considered, and a win-win situation of a power grid and users is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power system operation and dispatch, and in particular to a multi-substation area charging dispatch method and system considering the changing trend of charging load. Background Technique

[0002] The statements in this part only provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] For the power grid, electric vehicles during charging are a kind of load. When there are a large number of charging electric vehicles in a certain area of the substation area, the formed load fluctuation impacts the power grid. At the same time, the load formed by electric vehicles has randomness, indirectly causing random fluctuations in the load and resulting in load imbalance. The unreasonable distribution of the load will lead to a decline in the voltage quality in the substation area. When a certain area of the substation area is overloaded, the line voltage drop increases, which may cause the voltage at the user end to be lower than the normal range.

[0004] Traditional single-substation area optimization can limitedly reduce the pressure of the load on the substation area, but there are deficiencies in power supply reliability, resource utilization rate, and adaptability to load growth. The optimization of multiple substation areas requires the collaborative optimization of multiple objectives, and traditional optimization algorithms cannot meet the requirements of multiple objectives for the optimization of multiple substation areas.

[0005] The charging load data of the substation area is affected by factors such as emergencies, commissioning of new charging facilities, and evolution of residents' travel patterns. The data distribution often changes, and the adaptability of models with fixed structures and parameters is limited. It cannot effectively cope with the changes in the charging load characteristics of the substation area, and the prediction accuracy decreases over time, making it difficult to meet the requirements of precise power grid dispatch and intelligent management of charging facilities. Moreover, most of the substation area optimization strategies are single-substation area optimization strategies and do not involve the collaborative dispatch of multiple substation areas, resulting in unsatisfactory optimization effects. Summary of the Invention

[0006] In order to solve the technical problems existing in the above background technique, the present invention provides a multi-substation area charging dispatch method and system considering the changing trend of charging load, optimizes the dispatch according to the dynamic characteristics of the load, and realizes a win-win situation for the power grid and users by considering the collaborative optimization of multiple substation areas.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] The first aspect of the present invention provides a multi-substation area charging dispatch method considering the changing trend of charging load, including the following steps:

[0009] Preprocess the historical charging load data of the substation area, and use the data of a set time period as a sample;

[0010] Train a neural network model based on the obtained samples, and combine with the currently collected load data to obtain a load prediction curve;

[0011] Based on the obtained load prediction curve, as well as the status information of load transfer in each substation area, the load information of the substation area, the change rate per unit time, the voltage information, and the load of the tie line, with the goals of load balancing, minimizing network loss, and maximizing the utilization rate of charging piles, and with active power imbalance, reactive power imbalance, and line overload as constraints, use deep reinforcement learning to obtain an optimization strategy for multiple substation areas to achieve load balance.

[0012] As a further implementation method, the preprocessing includes, by collecting the historical charging load data of the substation area, using the kernel density estimation method to convert the load in each set time period into a kernel density estimation map, and dividing it into time windows.

[0013] As a further implementation method, the preprocessing also includes, by comparing the Jensen-Shannon divergence of the data metric time series data in each set time period to determine the change in the distribution followed, and determining the distance between samples.

[0014] As a further implementation method, train a neural network model based on the obtained samples, and combine with the currently collected load data to obtain a load prediction curve; including: using the trained long short-term memory network to check the P value of the observed divergence value related to the historical divergence value distribution. If it exceeds the significance level τ, the behavior of the consumption load has not changed; otherwise, it has changed; when it changes, retrain the model and remember the previous weights.

[0015] As a further implementation method, use deep reinforcement learning to obtain an optimization strategy for multiple substation areas, including: by learning the action value function, estimating the expected long-term return of taking a certain action in a given state.

[0016] As a further implementation method, using deep reinforcement learning to obtain an optimization strategy for multiple substation areas also includes: using various variables in the state space to describe the operating conditions of the power system, using the action space to describe the range of interaction and change between substation areas, and setting a reward function to make the deep reinforcement learning network converge to the optimal direction.

[0017] As a further implementation method, the state space includes the load of the current substation area, the original load ratio transferred from substation area j to substation area i, the total load power of substation area i, the change rate of substation area i per unit time, the impact of substation area i on the voltage of substation area j, and the load rate of the tie line;

[0018] The action space includes the capacity of tie line ij, the load of tie line ij before load transfer, and the transferred load;

[0019] The reward function includes load balancing, minimizing network loss, and maximizing the utilization rate of charging piles.

[0020] The second aspect of the present invention provides a multi - sub - area charging scheduling system that considers the changing trend of charging load, including:

[0021] A data acquisition and pre - processing module, configured to: pre - process the historical charging load data of the sub - area, and use the data within a set time period as a sample;

[0022] A load prediction module, configured to: train a neural network model based on the obtained sample, and combine the currently collected load data to obtain a load prediction curve;

[0023] A multi - sub - area optimization scheduling module, configured to: based on the obtained load prediction curve, as well as the status information of load transfer in each sub - area, sub - area load information, change rate per unit time, voltage information, and tie - line load, with the goals of load balancing, minimizing network loss, and maximizing the utilization rate of charging piles, and with active power imbalance, reactive power imbalance, and line overload as constraints, use deep reinforcement learning to obtain a multi - sub - area optimization strategy to achieve load balancing.

[0024] The third aspect of the present invention provides a computer - readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps in the above - mentioned multi - sub - area charging scheduling method based on considering the changing trend of charging load are implemented.

[0025] The fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps in the above - mentioned multi - sub - area charging scheduling method based on considering the changing trend of charging load are implemented.

[0026] Compared with the prior art, the above - mentioned one or more technical solutions have the following beneficial effects:

[0027] By pre - processing the historical load data and the currently collected data, analyzing the load of each sub - area, with the goals of load balancing, minimizing network loss, and maximizing the utilization rate of charging piles, and with active power imbalance, reactive power imbalance, and line overload as constraints, using deep reinforcement network to learn the optimal strategy, obtain the optimal action, and finally coordinate the sub - area load according to the optimal action to complete the balancing of the sub - area load. During this period, optimize the scheduling according to the dynamic characteristics of the load, consider the collaborative optimization of multiple sub - areas, and achieve a win - win situation for the power grid and users. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The specification drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0029] Figure 1It is a schematic diagram of the multi-substation area charging scheduling process considering the changing trend of charging load provided by one or more embodiments of the present invention;

[0030] Figure 2 It is a schematic diagram of the optimization strategy considering multiple substation areas using deep reinforcement learning provided by one or more embodiments of the present invention. Detailed implementation manners

[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0032] It should be noted that the following detailed descriptions are all exemplary and are intended to provide further explanations of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0033] It should be noted that the terms here are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should also be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or their combinations.

[0034] As introduced in the background technology, the charging load data of the substation area is affected by factors such as emergencies, commissioning of new charging facilities, and evolution of residents' travel patterns. The data distribution often changes, and the adaptability of models with fixed structures and parameters is limited. It cannot effectively cope with the changes in the charging load characteristics of the substation area, the prediction accuracy decreases over time, and it is difficult to meet the needs of precise grid scheduling and intelligent management of charging facilities. Moreover, most of the optimization strategies for the substation area are single-substation area optimization strategies and do not involve the optimization and coordinated scheduling of multiple substation areas, resulting in unsatisfactory optimization effects.

[0035] In some existing technologies, the orderly charging scheduling method based on the load prediction of the deep learning algorithm has become increasingly mature. For example, "A Method for Orderly Charging of Electric Vehicles in the Substation Area Based on Load Prediction and Deep Reinforcement Learning" (CN114169593A) predicts the future remaining charging capacity using a fully connected feedforward neural network (FFN) according to the historical remaining charging capacity information of the substation.

[0036] For example, "A Method for Orderly Charging Strategy of Charging Vehicles" (CN116957845A) trains through the user vehicle charging data collected by the data training module and obtains the training result of the user charging data. The data analysis module conducts user charging analysis based on the training result of the user charging data; the data recommendation module recommends the user vehicle charging strategy based on the user charging analysis result.

[0037] These existing technologies are all single-substation area optimization strategies and do not involve the optimized collaborative scheduling of multiple substation areas, resulting in unsatisfactory optimization effects.

[0038] Therefore, the following embodiments provide a multi-substation area charging scheduling method and system considering the changing trend of charging load, which optimize the scheduling according to the dynamic characteristics of the load and consider the collaborative optimization of multiple substation areas to achieve a win-win situation for the power grid and users.

[0039] Embodiment 1:

[0040] A multi-substation area charging scheduling method considering the changing trend of charging load includes the following steps:

[0041] S1 Charge load data collection and analysis. By collecting the historical charging load data of the substation area, the load of each day is transformed into a kernel density estimation map using KDE (Kernel Density Estimation) for the data of each day, and the data of each day is divided into time windows.

[0042] S2 Compare the daily data metrics using JSD divergence to determine the distance between samples by comparing the distribution changes of time series data on each day.

[0043] S3 Construct a day-ahead load prediction model. Use the data to train an Adaptive Long Short-Term Memory Network (DA-LSTM). This model can dynamically detect change points without a fixed detection threshold, retrain the model when the data exceeds the threshold, and remember the previous weights.

[0044] S4 Multi-substation area optimization ideas and goals. In a multi-substation area system, according to the obtained load prediction curve, analyze the load of each substation area. With the goal of load balancing, minimizing network loss, and maximizing the utilization rate of charging piles, and with the constraints of active power imbalance, reactive power imbalance, and line overload, use deep reinforcement learning to solve the joint collaborative scheduling of multiple substation areas and achieve balanced load.

[0045] The multi-substation area optimization strategy based on the changing trend of charging load given in this embodiment, through the collection and analysis of data, calculates the JSD divergence, constructs a load prediction model, analyzes the data characteristics, the information collection unit collects the status information of the substation area, establishes an interactive data set, analyzes the load of each substation area. With the goal of load balancing, minimizing network loss, and maximizing the utilization rate of charging piles, and with the constraints of active power imbalance, reactive power imbalance, and line overload, uses a deep reinforcement network to learn the optimal strategy, obtains the optimal action, and finally coordinates the load of the substation area according to the optimal action to complete the balancing of the load of the substation area.

[0046] The multi-substation area optimization strategy considering the changing trend of charging load proposed in this embodiment, as Figure 1 shown, includes:

[0047] S1 Charge load data collection and analysis. Collect historical charging load data of the substation area through smart meters, obtain historical load data with a set step size (for example, the step size is 2 hours), and use KDE (Kernel Density Estimation) to convert the daily load into a kernel density estimation map for the data of each day, and divide the daily data into time windows.

[0048] As a further implementation, KDE (Kernel Density Estimation) places a Gaussian kernel at each data point and then adds these kernels together to obtain the final density estimate. This method can form larger values where the data point density is high and lower values where there are few data points. The Gaussian kernel is as shown in the following formula:

[0049]

[0050] In the formula, h is the bandwidth, y is the target value, and Y i is the sample value.

[0051] S2 Compare the daily data metric time series data to determine the distance between samples by using the JSD divergence to compare the distribution changes followed by the time series data each day. The JSD divergence can measure the magnitude of the distribution changes followed by the time series data at a specific time point.

[0052] S3 Construction of the day-ahead load forecasting model. Use the data to train the Adaptive Long Short-Term Memory Network (DA-LSTM).

[0053] The Long Short-Term Memory Network (DA-LSTM) can dynamically check for change points without fixing the detection threshold, retrain the model when the data exceeds the threshold, and remember the previous weights.

[0054] As a further implementation, the network checks the P value of the observed divergence value related to the historical divergence value distribution. If the value exceeds the significance level, the algorithm considers that the behavior of the consumption load has not changed.

[0055] As a further implementation, the P value is equal to the area at both ends under the probability density function curve. In this case, the distribution of the divergence value changes with more observation records of the load, as shown in the following formula:

[0056] P = 2×(1 - Φ(|z|));

[0057]

[0058] where Φ(z) represents the probability that the random variable is less than or equal to z.

[0059] As a further implementation, the significance level τ represents how extreme the divergence can be before a change signal is sent.

[0060] The main advantage of using the P-value is that it is more persistent than the threshold, which may be unstable in the change detection problem. The significance level τ detects the sensitivity of changes, regardless of whether the magnitude of such changes is measured by the divergence metric.

[0061] In this embodiment, training an Adaptive Long Short-Term Memory Network (DA-LSTM) using data may include the following steps:

[0062] Dynamic model initialization: Construct the DA-LSTM network architecture, initialize the model weights; load the initial historical load data for pre-training; establish the initial divergence value distribution benchmark (probability density function).

[0063] Continuous prediction and monitoring: The model receives new load observation data in real time; generates prediction values and calculates prediction errors; converts the divergence value (prediction error metric) of the current observation window into a statistic.

[0064] Adaptive distribution update: Maintain a dynamically updated historical divergence value distribution library; adopt a sliding window mechanism to update the probability density function; automatically adjust the distribution parameters (such as mean and variance) as the data volume grows.

[0065] Hypothesis testing mechanism: Calculate the P-value corresponding to the current divergence value; compare the calculated P-value with the preset significance level τ.

[0066] Change point decision logic:

[0067] When P ≤ τ: Trigger a change point alarm, retain the current model weights as initialization parameters, and start the model re-training process;

[0068] When P > τ: Determine that the system behavior has not changed significantly, and continue to accumulate new data to update the historical distribution.

[0069] Incremental model update: Adopt a transfer learning strategy for model re-training; freeze some of the underlying network layers and fine-tune the top-level structure; apply weight regularization techniques to retain historical knowledge; the updated model inherits the historical weights and continues to monitor.

[0070] DA-LSTM has a dual adaptation mechanism: It includes both the dynamic adaptation of the LSTM network parameters and the autonomous evolution of the statistical detection threshold. It also has a memory retention feature: It retains the memory of historical patterns through the weight inheritance strategy to avoid catastrophic forgetting.

[0071] At the same time, the sensitivity is adjustable: By adjusting the τ value, the detection sensitivity and false alarm rate are balanced. This solution combines deep learning with statistical process control to achieve intelligent detection of change points in non-stationary time series, which is particularly suitable for load prediction scenarios.

[0072] As a further implementation, Bayesian optimization is used to find the appropriate hyperparameters of DA-LSTM to achieve adaptive load forecasting in the power distribution area.

[0073] In this embodiment, Bayesian optimization is adopted. Since the selection of hyperparameters for deep learning networks often uses grid search or random search, but these are unconscious of past evaluations, resulting in slower search speeds and suboptimal hyperparameters. Bayesian optimization tracks previous evaluations by mapping hyperparameters to the probability of the target score through a surrogate model, and can find better hyperparameter combinations with fewer steps. At the same time, generally, the derivatives of the hyperparameters of neural networks are difficult to obtain, but Bayesian optimization does not require the calculation of derivatives. Therefore, in this embodiment, Bayesian optimization is used to obtain the hyperparameters of the deep learning network, and then the load forecasting curve is obtained.

[0074] S4 Multi-substation area optimization ideas and objectives

[0075] In a multi-substation area system, based on the load forecasting curve obtained in step S3, the load of each substation area is analyzed. With the goals of load balancing, minimizing network losses, and maximizing the utilization rate of charging piles, and with active power imbalance, reactive power imbalance, and line overload as constraints, deep reinforcement learning is used to solve the joint collaborative scheduling of multiple substation areas to achieve balanced load.

[0076] In reinforcement learning, the agent needs to interact with the environment to learn the optimal policy. DQN (Deep Q-Network) is an algorithm that combines deep learning and reinforcement learning and is used to estimate the action-value function (Q-function). The agent selects an action based on the current state. After executing the action, the environment will return a reward and enter the next state. In this embodiment, the "action" of reinforcement learning refers to the "instruction to transfer the power grid load".

[0077] The core of reinforcement learning is to learn a Q-function (action-value function), which is used to estimate the expected long-term return of taking a certain action in a given state. To train the neural network in DQN, a loss function with a penalty term is used to measure the gap between the predicted Q value and the "correct" Q value. This "correct" Q value is obtained through the target network or an estimate based on the reward and the next state.

[0078] In the problem of substation area load balancing, the state is the load situation, voltage and other information of each substation area, the action is the load transfer strategy, such as how much load to transfer from substation area A to substation area B, and the reward is the degree of achievement of optimization goals such as reducing the overload risk and reducing network losses. The specific method includes the following steps:

[0079] Step 1: The information collection unit collects the state information of load transfer in multiple substation areas, the load information of substation areas, the change rate per unit time, voltage information, and the load of tie lines.

[0080] Step 2: Establish the interaction sample data set of the controller's control actions and the distribution transformer area.

[0081] Step 3: Based on the interaction data set established in Step 2, adopt a deep reinforcement network to learn the optimal behavior strategy and input the optimal actions.

[0082] Step 4: Control the transfer of load in the distribution transformer area according to the optimal network learning strategy in Step 3.

[0083] The optimization objectives of the multi-distribution transformer area system are usually multi-faceted. It is necessary to simultaneously consider minimizing network losses, maximizing the utilization rate of charging piles, and some constraint conditions to achieve load balance in the distribution transformer area. Define the key elements of the deep Q network.

[0084] 1. State space (P i , P(ij), P(i), ΔP i , ΔV ij , L line ): In the context of multi-distribution transformer area collaboration, the state space refers to the set of all possible system states. It contains various variables that describe the operating conditions of the system, and these variables can provide sufficient information for the intelligent agent to make reasonable decisions.

[0085] As a further implementation method, P in the state space i represents the load of the current distribution transformer area i, P ij represents the original load ratio transferred from distribution transformer area j to distribution transformer area i, P(i) represents the total load power of distribution transformer area i, ΔP i represents the load change rate of distribution transformer area i per unit time, ΔV ij represents the influence of distribution transformer area i on the voltage of distribution transformer area j, and L line represents the load rate of the tie line.

[0086] 2. Action space: The action space contains the range of mutual interaction changes between distribution transformer areas, considering the capacity limitations of lines and transformers. The action space can be expressed as the following formula:

[0087] L ij +a ij ≤C ij ;

[0088] P i +a ij ≤P(i);

[0089] where C ij is the capacity of the tie line ij, L ij represents the load of the tie line ij before transferring the load, and a ij is the transferred load.

[0090] 3. Reward: Use the reward function to make the network converge in the optimal direction.

[0091] (1) Load standard deviation: The standard deviation is a statistic that measures the degree of data dispersion. By minimizing the standard deviation of the load rate, the control strategy can be motivated to take actions to make the load rates of each substation area closer to the average value, thereby achieving load balancing, as shown in the following formula:

[0092]

[0093] Among them, represents the mean value of the load rates of all substation areas.

[0094] (2) Minimizing power losses: It can guide more reasonable load transfer during load transfer to achieve the purpose of energy conservation. Power losses mainly include line losses and transformer losses. The transformer losses can be mainly calculated based on the parameters and load conditions of the transformer, as shown in the following formula:

[0095]

[0096] Among them, P loss represents the total power loss, represents the line losses between substation areas, represents the transformer losses of the substation area, R ij represents the line resistance, I ij represents the line current.

[0097] (3) Charging pile utilization rate: Improving the charging pile utilization rate is an important goal, which can prompt the intelligent agent to take measures to reasonably guide electric vehicles to charge in the substation area with more idle charging piles, or adjust the charging power of the charging piles to improve the overall charging efficiency, so as to better meet the charging needs of users. As shown in the following formula:

[0098]

[0099] Among them, represents the number of used charging piles in the substation area, represents the total number of charging piles, U i represents the utilization rate of the charging pile in the i-th substation area, and U represents the utilization rate of the entire charging pile.

[0100] Comprehensive reward: The comprehensive reward function can balance the relationship between multiple optimization goals. As shown in the following formula:

[0101] r = w 1 r 1 + w 2 r 2 + w 3 r 3 ;

[0102] Among them, w 1 ~w 3represents the discount factor, r 1 ~r 3 represents the optimization objective.

[0103] The loss function considering the penalty term for constraint violation. The penalty term in the loss function takes effect during the training phase of the neural network. It directly affects the update of network parameters, making the network pay more attention to not violating the constraint conditions during the learning process. When dealing with line capacity constraints, the reward function may guide the learning by gradually reducing the reward for actions that cause line overload, while the penalty term in the loss function will immediately increase the loss every time line overload occurs, enabling the network to avoid generating such infeasible actions more quickly.

[0104] C = (ΔP, ΔQ, O);

[0105] where ΔP represents the imbalance of active power, ΔQ represents the imbalance of reactive power, and O represents line overload.

[0106] The Bellman optimal action-value function. In the optimal case, the optimal value of the current state-action (s, a) pair is equal to the reward obtained currently plus the maximum expected discounted reward that can be obtained by acting according to the optimal policy in the future state. As shown in the following formula:

[0107]

[0108] where Q * (s, a) represents the optimal action-value function for performing action a in state s, r(s, a) is the reward obtained immediately after performing the action in the state, γ is the discount factor, is the maximum optimal action value among all possible actions a' in the subsequent state s', representing the maximum expected discounted reward that can be obtained in the future state according to the optimal policy.

[0109] The loss function, as shown in the following formula:

[0110]

[0111] where μ represents the discount factor, Q(s, a; θ) represents the predicted value, y represents the target value, and c represents the penalty term.

[0112] Using deep reinforcement learning to consider the multi-substation area optimization strategy, this strategy includes, as Figure 2 shown, the agent selects an action according to the current state. After performing the action, the environment will return a reward and enter the next state. Its core is to learn a Q-function (action-value function), which is used to estimate the expected long-term return of taking a certain action in a given state. By using a loss function to measure the gap between the predicted Q value and the "correct" Q value, the optimal action is obtained, and the neural network parameters in the DQN are trained.

[0113] This application presents a multi-substation optimization strategy based on the changing trend of charging load. By collecting and analyzing data, calculating the Jensen-Shannon divergence (JSD), constructing a load prediction model, analyzing data characteristics, the information acquisition unit collects the status information of the substations, establishes an interactive data set, analyzes the load of each substation, aims at load balancing, minimizing network loss, and maximizing the utilization rate of charging piles, with the imbalance of active power, the imbalance of reactive power, and line overload as constraints, uses a deep reinforcement network to learn the optimal strategy, obtains the optimal action, and finally coordinates the substation load according to the optimal action to complete the balancing of the substation load. During this period, the scheduling is optimized according to the dynamic characteristics of the load, and the coordinated optimization of multiple substations is considered to achieve a win-win situation for the power grid and users.

[0114] By collecting and analyzing data, calculating the Jensen-Shannon divergence (JSD), constructing a load prediction model, analyzing data characteristics, the information acquisition unit collects the status information of the substations, establishes an interactive data set, analyzes the load of each substation, aims at load balancing, minimizing network loss, and maximizing the utilization rate of charging piles, with the imbalance of active power, the imbalance of reactive power, and line overload as constraints, uses a deep reinforcement network to learn the optimal strategy, obtains the optimal action, and finally coordinates the substation load according to the optimal action to complete the balancing of the substation load. During this period, the scheduling is optimized according to the dynamic characteristics of the load, and the coordinated optimization of multiple substations is considered to achieve a win-win situation for the power grid and users.

[0115] Through the coordinated optimization of multiple substations, the problem of load imbalance that cannot be addressed by single-substation optimization is solved, and the stability and reliability of the power grid are improved.

[0116] At the same time, through the analysis of the changing trend of charging load, the scheduling strategy is dynamically adjusted to adapt to the time-varying characteristics of the charging load, further improving the operating efficiency of the power grid. In terms of technical feasibility, kernel density estimation (KDE), Jensen-Shannon divergence (JSD), adaptive long short-term memory network (DA-LSTM), and deep reinforcement learning (DQN) are all mature technologies, which are easy to implement and apply.

[0117] In addition, through load balancing and minimizing network loss, the operating cost of the power grid is reduced; by maximizing the utilization rate of charging piles, the usage efficiency of charging facilities is improved, with significant economic benefits.

[0118] In terms of social benefits, the stability and reliability of the power grid are improved, the voltage fluctuation at the user end is reduced, the user satisfaction is improved, and at the same time, the popularization and application of electric vehicles are promoted, and the development of green energy is driven.

[0119] Embodiment 2:

[0120] A multi-substation charging scheduling system considering the changing trend of charging load, including:

[0121] A data acquisition and preprocessing module, configured to: preprocess the historical charging load data of the power distribution area, and use the data within a set time period as a sample;

[0122] A load prediction module, configured to: train a neural network model based on the obtained sample, and combine the currently collected load data to obtain a load prediction curve;

[0123] A multi-power-distribution-area optimization scheduling module, configured to: based on the obtained load prediction curve, as well as the state information of load transfer, load information of each power distribution area, change rate per unit time, voltage information, and tie line load of each power distribution area, with the goals of load balancing, minimizing network loss, and maximizing the utilization rate of charging piles, and with active power imbalance, reactive power imbalance, and line overload as constraints, use deep reinforcement learning to obtain a multi-power-distribution-area optimization strategy to achieve load balancing.

[0124] By preprocessing the historical load data and the currently collected data, analyzing the load of each power distribution area, with the goals of load balancing, minimizing network loss, and maximizing the utilization rate of charging piles, and with active power imbalance, reactive power imbalance, and line overload as constraints, using deep reinforcement network to learn the optimal strategy to obtain the optimal action, and finally coordinating the load of the power distribution area according to the optimal action to complete the balancing of the load of the power distribution area. During this period, optimize the scheduling according to the dynamic characteristics of the load, consider the collaborative optimization of multiple power distribution areas, and achieve a win-win situation for the power grid and users.

[0125] Embodiment 3:

[0126] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps in the multi-power-distribution-area charging scheduling method based on considering the changing trend of charging load as described in Embodiment 2 above.

[0127] Embodiment 4:

[0128] This embodiment provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the multi-power-distribution-area charging scheduling method based on considering the changing trend of charging load as described in Embodiment 2 above.

[0129] The steps involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation manners, reference may be made to the relevant description part of Embodiment 1. The term "computer-readable storage medium" should be understood to include a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0130] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A multi-zone charging scheduling method considering the changing trend of charging load, characterized in that: The following steps are involved: Pre-process the historical charging load data of the substation, using the data of the set time period as samples; The neural network model is trained based on the obtained samples and combined with the currently collected load data to obtain the load prediction curve; According to the obtained load forecast curve, as well as the status information of load transfer in each substation, substation load information, rate of change per unit time, voltage information and interconnection line load, with the goals of load balancing, minimizing network losses and maximizing charging pile utilization, and with active power imbalance, reactive power imbalance and line overload as constraints, deep reinforcement learning is used to obtain a multi-substation optimization strategy to achieve load balancing.

2. The multi-station charging scheduling method considering the charging load change trend as claimed in claim 1 is characterized in that: The preprocessing includes collecting the historical charging load data of the substation area, using the kernel density estimation method, converting the load of each set time period into a kernel density estimation graph, and dividing it into time windows.

3. The multi-station charging scheduling method considering the charging load change trend as claimed in claim 1, characterized in that: Preprocessing also includes comparing the distribution changes followed by the data measurement time series data in each set time period through JSD divergence to determine the distance between samples.

4. The multi-station charging scheduling method considering the charging load change trend as claimed in claim 1 is characterized in that: The neural network model is trained according to the obtained samples, and the load forecast curve is obtained by combining the currently collected load data; including: using the trained long short-term memory network, checking the P value of the observed divergence value related to the historical divergence value distribution, if it exceeds the significance level τ, the behavior of the consumption load has not changed; otherwise, it has changed; retraining the model when changes occur, and remembering the previous weights.

5. The multi-zone charging scheduling method considering the charging load change trend as claimed in claim 1, characterized in that: Use deep reinforcement learning to obtain multi-zone optimization strategies, including: estimating the expected long-term reward of taking an action in a given state by learning the action-value function.

6. The multi-zone charging scheduling method considering the charging load change trend as claimed in claim 1 is characterized in that: The multi-zone optimization strategy is obtained by using deep reinforcement learning, which also includes: using state space to describe various variables of the power system operating conditions, using action space to describe the range of mutual intersection and change of zones, and setting a reward function to make the deep reinforcement learning network converge to the optimal direction.

7. The multi-station charging scheduling method considering the charging load change trend as claimed in claim 6, characterized in that: The state space includes the load of the current substation, the original load ratio of substation j transferred to substation i, the total load power of substation i, the rate of change of substation i per unit time, the impact of substation i on the voltage of substation j, and the load rate of the tie line; The action space includes the capacity of tie line ij, the load of tie line ij before load transfer, and the transferred load; The reward functions include load balancing, minimizing network losses and maximizing charging pile utilization.

8. A multi-zone charging dispatching system considering the changing trend of charging load, characterized in that: include: The data acquisition and preprocessing module is configured to: preprocess the historical charging load data of the substation, using the data of a set time period as a sample; The load forecasting module is configured to: train a neural network model according to the obtained samples, and obtain a load forecasting curve in combination with the currently collected load data; The multi-station optimization scheduling module is configured as follows: based on the obtained load forecast curve, the status information of the load transfer in each station, the station load information, the rate of change per unit time, the voltage information and the interconnection line load, with the goals of load balancing, minimizing network losses and maximizing the utilization rate of charging piles, and with active power imbalance, reactive power imbalance and line overload as constraints, deep reinforcement learning is used to obtain the multi-station optimization strategy to achieve balanced load.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the steps in the multi-station area charging scheduling method considering the charging load change trend as described in any one of claims 1 to 7 are implemented.

10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the multi-station area charging scheduling method considering the charging load change trend as described in any one of claims 1 to 7 are implemented.

Images