Multi-time scale matching-based light-vehicle load cooperative scheduling method

Abstract

The invention relates to the technical field of intelligent power grid dispatching, and provides a multi-time scale matching-based light-vehicle load collaborative dispatching method, which comprises the following steps of: performing run-length analysis on data to extract run-length characteristics by synchronously acquiring photovoltaic output, electric vehicle charging power and power grid node voltage data; constructing a space-time diagram of dynamic node attributes and edge weights by using run length features, and inputting a prediction model to output multi-time scale probability prediction; mapping the run length characteristics into dynamic security constraints, and generating a power instruction sequence by adopting double-layer model prediction control rolling optimization in combination with a prediction result; and after the instruction is locally and adaptively executed, the mapping relation and the model parameters are optimized through closed-loop feedback data. According to the method, the problem of time sequence mismatch between the prediction model and the real-time control instruction is solved, and the dynamic response performance of the system to random fluctuation is improved.

Description

Technical Field

[0001] This invention relates to the field of smart grid dispatching technology, and particularly to a method for coordinated dispatching of optical vehicles and loads based on multi-timescale matching. Background Technology

[0002] Existing smart grid multi-timescale collaborative scheduling technologies for photovoltaics, vehicles, and loads suffer from the following technical challenges: Specifically, the multi-timescale characteristics of photovoltaic output, electric vehicle charging behavior, and load demand in smart grid collaborative scheduling lead to discrepancies between the update cycle of prediction models and the generation timing of real-time control commands. Prediction models are typically trained and updated based on historical data over long time spans, providing power predictions for the next few hours. However, real-time control commands need to be adjusted based on the actual system state at the second or minute level. This timing mismatch prevents control commands from effectively utilizing the latest prediction information. For example, when photovoltaic power generation fluctuates due to rapid cloud changes, scheduling strategies based on outdated predictions may fail to adjust electric vehicle charging power in a timely manner to smooth out power fluctuations, resulting in system frequency deviations or voltage overruns and a decline in dynamic response performance. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a vehicle-optical-load collaborative scheduling method based on multi-timescale matching, which solves the technical problem of degraded system dynamic response performance caused by the timing mismatch between prediction models and real-time control commands at multiple timescales.

[0004] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0005] This invention provides a method for coordinated scheduling of optical vehicles and payloads based on multi-timescale matching, comprising:

[0006] Step 1: Perform data acquisition to obtain synchronous measurement data of photovoltaic power output, electric vehicle charging power, and grid node voltage;

[0007] Step 2: Perform run analysis on the photovoltaic output data or grid node voltage data in the synchronous measurement data to extract run features, including run length distribution and run intensity;

[0008] Step 3: Using the extracted run features, generate feature vectors describing the dynamic attributes of the power grid topology nodes. Combine the power grid topology connection relationship with real-time power exchange data to construct a spatiotemporal graph with dynamic node attributes and edge weights.

[0009] Step 4: Input the spatiotemporal diagram into the prediction model and output the photovoltaic power output prediction, load demand prediction and electric vehicle dispatchable power prediction at different time scales in the future.

[0010] Step 5: Map the extracted run features to dynamic safety constraints and construct a model predictive control rolling optimization model. The rolling optimization model takes the photovoltaic output prediction, load demand prediction, electric vehicle dispatchable power prediction and the dynamic safety constraints as inputs and solves the model in a rolling manner to generate power command sequences for photovoltaic inverters, electric vehicle charging piles and controllable loads.

[0011] Step 6: Issue and execute the power command sequence; collect the actual power and voltage data of the system after executing the power command sequence; use the collected actual power and voltage data of the system and the run features extracted by the run analysis to adjust the mapping relationship between the run features and the dynamic safety constraints, and adjust the parameters of the rolling optimization model.

[0012] Furthermore, in the optical vehicle-load cooperative scheduling method based on multi-timescale matching described in this invention, step 2 includes:

[0013] For the photovoltaic power output data or grid node voltage data in the synchronous measurement data, perform adaptive noise complete empirical mode decomposition to obtain multiple intrinsic mode function components;

[0014] For each obtained intrinsic mode function component, the time series mean of the component is calculated, and the mean is used as a threshold.

[0015] In the time series of the component, segments that are continuously above or continuously below the threshold are detected, and each segment is marked as a run;

[0016] The number of data points included in each run is counted as the run length; the sum of squares of the deviations of the amplitudes of all data points in each run from the threshold is calculated as the run intensity.

[0017] Based on the run length and run intensity of all runs within a time window, the run length distribution, average run intensity, and run information entropy are determined. The run characteristics include the run length distribution, average run intensity, and run information entropy.

[0018] Furthermore, in the optical vehicle-load cooperative scheduling method based on multi-timescale matching described in this invention, step 3 includes:

[0019] The run features extracted from the synchronous measurement data are collected, and normalization and principal component analysis are performed on the collected run features to generate the feature vector of each monitoring node in the corresponding power grid topology.

[0020] Obtain the topological connections of the power grid, and construct a static topology graph with electrical nodes as vertices and electrical connections as edges;

[0021] Calculate the power flow values ​​between nodes based on the real-time power exchange data in the synchronous measurement data;

[0022] The weights of the corresponding edges in the static topology graph are dynamically adjusted using the calculated power flow values.

[0023] The generated feature vectors are assigned as node attributes to the corresponding vertices in the static topology graph to obtain the spatiotemporal graph.

[0024] Furthermore, in the optical vehicle-load cooperative scheduling method based on multi-timescale matching described in this invention, step 4 includes:

[0025] Based on the attributes of vertices and the weights of edges in the spatiotemporal graph, the attention coefficient between a vertex and its neighboring vertices is calculated.

[0026] The attention coefficients are used to weight the attributes of neighboring vertices and update the feature representation of the vertices.

[0027] The updated vertex feature sequence is input into three processing branches: long-term, medium-term, and short-term. In each branch, a temporal convolutional network is applied to process the feature sequence. The dilation coefficients of the temporal convolutional networks in the long-term, medium-term, and short-term branches are different.

[0028] Cross-attention calculation is performed between the long-term processing branch and the short-term processing branch to fuse the features output by the long-term branch and the features output by the short-term branch.

[0029] The merged features are decoded to output the probability distribution of photovoltaic power output, load demand, and dispatchable power of electric vehicles at different time scales in the future.

[0030] Furthermore, in the optical vehicle-load cooperative scheduling method based on multi-timescale matching described in this invention, step 5 includes:

[0031] A pre-defined mapping rule base is provided, which defines the mapping relationship between run-length characteristics and system safe operation boundaries.

[0032] The extracted run-length features are matched with the run-length features defined in the mapping rule base;

[0033] When a match is successful, at least one constraint is generated based on the mapping relationship, and the dynamic security constraint includes the generated constraint.

[0034] Furthermore, in the optical vehicle-load cooperative scheduling method based on multi-timescale matching described in this invention, step 5 further includes:

[0035] The upper-level optimization calculation is performed in the first cycle. The input of the upper-level optimization calculation includes the long-term scale portion of the photovoltaic output forecast, load demand forecast and electric vehicle dispatchable power forecast, as well as the dynamic security constraints triggered by the long run characteristics in the run length distribution. The upper-level optimization calculation aims at optimal economic efficiency and satisfies the physical constraints of the power grid, and outputs the baseline power plan for future periods.

[0036] The lower-level rolling optimization calculation is performed in a second cycle, which is shorter than the first cycle. The inputs to the lower-level rolling optimization calculation include the short-to-medium-term scale portions of the photovoltaic output forecast, load demand forecast, and electric vehicle dispatchable power forecast, the real-time power data in the synchronous measurement data, the baseline power plan, and dynamic safety constraints triggered by the high run intensity characteristics in the run intensity.

[0037] Within each second cycle, a real-time optimization problem is constructed. The objective of the real-time optimization problem is to minimize the deviation between the actual power and the baseline power plan and to suppress power fluctuations. The constraints include grid physical constraints and dynamic security constraints triggered by the high run-length strength characteristics of the current input. The real-time optimization problem is solved to obtain the power command sequence.

[0038] Furthermore, in the optical vehicle-load cooperative scheduling method based on multi-timescale matching described in this invention, solving the real-time optimization problem to obtain the power command sequence includes:

[0039] To address the uncertainty parameters introduced by prediction in the real-time optimization problem, a split-Brow bar optimization method is used.

[0040] Construct a fuzzy set, which is centered on the probability distribution of the corresponding short-to-medium-term scale in the photovoltaic power output prediction, load demand prediction and electric vehicle dispatchable power prediction;

[0041] Based on the fuzzy set, the objective function of the real-time optimization problem is reconstructed as minimizing the system performance deviation and fluctuation in the worst case.

[0042] By applying the alternating direction multiplier method, the reconstructed real-time optimization problem is decomposed into multiple subproblems, each subproblem corresponding to a type of adjustable resource;

[0043] The multiple sub-problems are computed in parallel, and information is exchanged through global coordination variables during the computation process. The solutions to the sub-problems are iteratively updated until the solutions converge. The converged solutions are then used as the power command sequence.

[0044] Furthermore, in the optical vehicle-load cooperative scheduling method based on multi-timescale matching described in this invention, step 6 includes:

[0045] Issue the power command sequence;

[0046] At the command execution end, the power command sequence is received and the target power value is extracted.

[0047] Acquire local electrical quantity measurements at a rate higher than the power command sequence update frequency;

[0048] A first-order hysteresis filtering algorithm is used to smooth the target power value to obtain a smoothed power command.

[0049] The actual output power is calculated based on the collected local electrical quantity measurements.

[0050] Calculate the deviation between the actual output power and the smoothed power command, and apply a proportional-integral control algorithm to the deviation to generate a power compensation amount;

[0051] The power compensation amount is superimposed on the smoothed power command to obtain the final executed power command.

[0052] Furthermore, in the optical vehicle-load cooperative scheduling method based on multi-timescale matching described in this invention, step 6 further includes:

[0053] The collected actual system power and voltage data are aligned with the run-length features extracted during the execution of the power command sequence by timestamps;

[0054] Based on the aligned data, the execution effect of the power command sequence under different run characteristics is evaluated, and a reward signal is generated;

[0055] Define a state vector, which includes a run feature vector and key system state variables; define an action vector, which includes adjustments to parameters in the mapping rule base or adjustments to the weights of the objective function of the rolling optimization model.

[0056] Using the generated reward signal, a policy function is trained through a reinforcement learning algorithm, with the training objective being to maximize the expected value of future cumulative rewards.

[0057] Using the trained policy function, an action vector is output based on the current state vector. The parameters of the mapping rule base and the rolling optimization model are then adjusted based on the output action vector.

[0058] Furthermore, in the optical vehicle-load cooperative scheduling method based on multi-timescale matching described in this invention, the training of the policy function through a reinforcement learning algorithm includes:

[0059] Initialize the parameters of the deep Q-network, which is used to evaluate the value of state-action pairs;

[0060] Initialize the experience replay buffer;

[0061] In each training iteration, the following operations are performed: a batch of experience tuples are sampled from the experience replay buffer. Each experience tuple includes a historical state vector, an action vector executed, a reward signal obtained, and a next state vector.

[0062] Input the sampled next state vector into the current depth Q network to calculate the maximum predicted action value for the next state;

[0063] Calculate the target value, which is the sum of the reward signal in the sampled empirical tuple and the discounted maximum predicted action value for the next state;

[0064] Input the historical state vector and the executed action vector from the sampled empirical tuple into the current depth Q network to obtain the current action value prediction;

[0065] The temporal difference error algorithm is used to calculate the loss between the target value and the current action value prediction.

[0066] The parameters of the current deep Q-network are updated using the calculated loss through backpropagation algorithm.

[0067] Periodically update the parameters of an independent target depth Q-network;

[0068] After training, the current deep Q-network is used as the policy function to select the action vector that maximizes the value of the predicted action based on the input state vector.

[0069] The beneficial effects of this invention are:

[0070] This invention extracts run features characterizing signal fluctuations from synchronous measurement data using run analysis technology, elevating existing time-series analysis to the level of fluctuation mode statistics, effectively quantifying the multi-timescale dynamic behavior of photovoltaic power output and grid node voltage. By constructing a spatiotemporal graph that integrates dynamic node attributes and real-time power weights, it addresses the insufficient spatial correlation characterization of existing prediction models, enabling prediction results to simultaneously reflect spatiotemporal evolution patterns. Dynamic constraints are generated using the mapping relationship between run features and system safety boundaries, directly embedding fluctuation characteristics into the optimization model, achieving time-series matching between prediction information and control commands. A two-layer model predictive control architecture is adopted: the upper layer optimizes to ensure economic efficiency, while the lower layer performs rolling optimization to handle real-time fluctuations. Reinforcement learning is used to continuously adjust mapping rules and model parameters, forming an adaptive closed-loop optimization system. Finally, a local adaptive execution strategy is used to mitigate command transmission delays, comprehensively improving the system's dynamic response capability to multi-timescale fluctuations. Attached Figure Description

[0071] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0072] Figure 1 This is a flowchart illustrating a collaborative scheduling method for optical vehicles and loads based on multi-timescale matching. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0074] To better understand the purpose of this invention, the invention will now be described in further detail.

[0075] This invention provides a method for coordinated scheduling of optical vehicles and payloads based on multi-timescale matching, comprising:

[0076] Step 1: Perform data acquisition to obtain synchronous measurement data of photovoltaic power output, electric vehicle charging power, and grid node voltage;

[0077] Step 2: Perform run analysis on the photovoltaic output data or grid node voltage data in the synchronous measurement data to extract run features, including run length distribution and run intensity;

[0078] Step 3: Using the extracted run features, generate feature vectors describing the dynamic attributes of the power grid topology nodes. Combine the power grid topology connection relationship with real-time power exchange data to construct a spatiotemporal graph with dynamic node attributes and edge weights.

[0079] Step 4: Input the spatiotemporal diagram into the prediction model and output the photovoltaic power output prediction, load demand prediction and electric vehicle dispatchable power prediction at different time scales in the future.

[0080] Step 5: Map the extracted run features to dynamic safety constraints and construct a model predictive control rolling optimization model. The rolling optimization model takes the photovoltaic output prediction, load demand prediction, electric vehicle dispatchable power prediction and the dynamic safety constraints as inputs and solves the model in a rolling manner to generate power command sequences for photovoltaic inverters, electric vehicle charging piles and controllable loads.

[0081] Step 6: Issue and execute the power command sequence; collect the actual power and voltage data of the system after executing the power command sequence; use the collected actual power and voltage data of the system and the run features extracted by the run analysis to adjust the mapping relationship between the run features and the dynamic safety constraints, and adjust the parameters of the rolling optimization model.

[0082] During the data acquisition phase, synchronous measurement devices deployed at the grid connection points of photovoltaic power plants, electric vehicle charging station bus lines, and key nodes of the distribution network synchronously acquire three-phase voltage and current waveform data at a sampling frequency of no less than 10kHz. The measurement devices achieve microsecond-level time synchronization through the IEEE 1588 precision clock protocol, assigning a unified time stamp to each data point. The acquired data is then processed in real-time to generate a sequence of instantaneous photovoltaic output values, an electric vehicle charging power sequence, and a sequence of effective node voltage values, forming a multi-source synchronous measurement dataset with strict time alignment.

[0083] An adaptive noise-complete empirical mode decomposition (EMD) process is performed on photovoltaic power output sequences or node voltage sequences. This algorithm decomposes the non-stationary signal into several intrinsic mode function (IMF) components and residual terms by repeatedly injecting adaptive white noise and calculating the mean envelope. For each IMF component, its time series mean is used as a threshold to identify sample intervals continuously above or below the threshold and label them as independent runs. The number of sample points in each run is counted as the run length, and the sum of the squares of the deviations of the sample amplitudes within a run from the threshold is calculated as the run intensity. Finally, based on the statistical results of all runs within the sliding time window, a feature set including a run length distribution histogram, average run intensity, and run information entropy is generated.

[0084] After normalizing the run-length features extracted from different monitoring points, principal component analysis (PCA) is used for dimensionality reduction and fusion to generate high-dimensional feature vectors representing the dynamic behavior of each power grid node. Simultaneously, a static topology graph is constructed based on the power grid wiring diagram, with electrical nodes as vertices and line connections as edges. The weights of the edges between nodes are then dynamically adjusted based on real-time power flow calculations. The feature vectors are then assigned as dynamic node attributes to the corresponding vertices of the topology graph, forming a spatiotemporal graph model where node attributes update with run-length features and edge weights change with power flow.

[0085] After the spatiotemporal graph is input into the prediction model, the spatial correlation strength between nodes is first calculated by a multi-head graph attention network. Neighbor node features are weighted and aggregated using attention coefficients to update the vertex feature representation. The updated feature sequence is then input into prediction branches at three time scales: long, medium, and short. Each branch uses a temporal convolutional network with different dilation coefficients to extract temporal features. The long-term branch focuses on trend components, while the short-term branch captures fluctuation details. A cross-attention mechanism is used to achieve multi-scale feature fusion. The final output layer generates probability distribution predictions for photovoltaic power output, load demand, and dispatchable power of electric vehicles for the next 15 minutes, 1 hour, and 4 hours.

[0086] The mapping from run characteristics to safety constraints is achieved through a pre-defined rule base. The rule base defines the correspondence between different combinations of run characteristics and system safety risks. For example, when the run intensity of the high-frequency component exceeds a threshold, constraints on the response rate of the energy storage system are automatically generated. The mapping process matches real-time run characteristics with the rule base, triggering the corresponding constraint generation logic and dynamically adjusting parameters such as power regulation rate and reserve capacity.

[0087] In the constructed two-layer model predictive control architecture, the upper-layer optimizer runs on an hourly cycle, integrating long-term forecast results and stable operating condition constraints to generate a baseline power plan with the goal of minimizing electricity purchase costs. The lower-layer optimizer executes on a minute-by-minute cycle, integrating short- and medium-term forecasts, real-time measurement data, and fluctuating operating condition constraints, and employs distributed bar optimization to handle forecast uncertainties. The optimization problem is decomposed into photovoltaic, charging pile, and load subproblems using the alternating direction multiplier method and solved in parallel, ultimately outputting the power command sequence for each execution unit.

[0088] After the command is issued, the execution end adopts a local adaptive control strategy. Based on the received power command, the charging pile controller combines millisecond-level sampled voltage and current data, smooths the command step through a first-order hysteresis filter, and uses a proportional-integral controller to compensate for execution deviations. This combination of global optimization and local fine-tuning effectively mitigates control errors caused by command transmission delays.

[0089] During the closed-loop optimization phase of the system, the actual power-voltage curves are spatiotemporally aligned with historical run characteristics to quantitatively evaluate the control effect under different fluctuation modes. A policy network is trained using deep reinforcement learning algorithms, with run characteristics and system state as inputs, dynamically adjusting constraint mapping rules and optimizing model weight parameters. This continuous learning mechanism based on actual operating data enables the system to gradually adapt to slowly time-varying factors such as changes in grid topology and the increasing penetration rate of new energy sources.

[0090] When performing adaptive noise-complete empirical mode decomposition (EMD), multiple intrinsic mode function (EMF) components are obtained by repeatedly adding white noise sequences with adaptive amplitude to the original signal and then performing EMD on the noisy signal. The results of multiple decompositions are then integrated and averaged to suppress noise interference. When calculating the time series mean of each EMF component, a sliding window is used to calculate the local mean to avoid the global mean being insufficiently adaptable to non-stationary signals. When detecting continuous segments, a minimum run length threshold is set to exclude pseudo-runs caused by transient noise. In the run strength calculation, a sum of squares method is used to highlight the contribution of large-amplitude fluctuations. The run information entropy is calculated based on the probability distribution of run length occurrences, and the Shannon entropy formula is used to quantify the randomness of the sequence.

[0091] The spatiotemporal graph construction method employs Z-score normalization to ensure comparability of run characteristics across different dimensions. Principal component analysis retains principal components with a cumulative contribution exceeding 85% to balance feature dimensionality and information completeness. The initial edge weights in the static topology graph are set to the reciprocal of the line impedance; during dynamic adjustment, weights are assigned signs based on real-time power flow direction—positive for forward flow and negative for reverse flow. The node attribute update cycle is synchronized with the run analysis window to guarantee temporal consistency of spatiotemporal evolution characteristics.

[0092] The prediction model implementation details include an additive attention mechanism for calculating the attention coefficient, mapping vertex attributes and edge weights to the attention score space using a trainable weight matrix. The dilation coefficient of the temporal convolutional network follows a geometric sequence, with long-term branches using an exponential growth pattern to expand the receptive field and short-term branches using a linear growth pattern to maintain temporal resolution. In the cross-attention mechanism, the feature vectors of short-term branches serve as query vectors, and the feature vectors of long-term branches serve as key-value pairs, achieving feature interaction through dot product attention. The probability distribution output layer uses a hybrid density network structure to simultaneously predict the mean and variance parameters of the power values.

[0093] The constraint mapping mechanism employs a production rule representation in its rule base. The antecedent of each rule is a logical combination of run-length features, and the consequent is a constraint generation instruction. The matching process uses a fuzzy matching algorithm, setting a feature similarity threshold to handle boundary cases. Constraint generation supports multiple types, including inequality constraints, equality constraints, and logical constraints. For example, when the run-length intensity of the high-frequency component of photovoltaic power output exceeds a threshold, an inequality constraint on the rate of change of electric vehicle charging power is generated. The effective duration of dynamic safety constraints is bound to the time window of the run-length features, ensuring timely matching between the constraints and the fluctuation characteristics.

[0094] The specific implementation of the described two-layer optimization model involves an upper-layer optimization calculation that integrates photovoltaic power output forecasts, load demand forecasts, and electric vehicle dispatchable power forecasts for the next four hours on an hourly cycle. It transforms the long run characteristics (durations exceeding a threshold) in the run length distribution into a relaxed constraint on the power regulation rate. The economic objective function includes grid purchase costs, network loss costs, and equipment operation and maintenance costs, which are solved using linear programming to obtain the baseline power plan for the next 24 hours. The lower-layer rolling optimization operates on a 15-minute cycle, fusing short-to-medium-term forecasts for the next hour with real-time power data. When the run intensity characteristic exceeds a threshold, it automatically generates strict constraints on the energy storage system's response time. The real-time optimization problem is modeled using quadratic programming, with the objective function including a plan tracking deviation term and a power fluctuation penalty term. The constraints integrate power flow equations, equipment capacity limits, and dynamic safety constraints.

[0095] The optimization process involves constructing a fuzzy set with the short-to-medium-term prediction probability distribution as the reference center, and setting a confidence interval to accommodate the prediction error distribution. The worst-case performance indicators are selected as the maximum system voltage deviation and the extreme value of frequency fluctuation as optimization objectives. The alternating direction multiplier method decomposes the global problem into a photovoltaic inverter reactive power regulation sub-problem, an electric vehicle charging power adjustment sub-problem, and a controllable load switching sub-problem. Each sub-problem is connected to the global coordination variable through Lagrange multipliers, and the augmented Lagrange method is used to handle coupling constraints during parallel solution. The convergence criterion is that both the original residual and the dual residual are simultaneously less than a set threshold; an adaptive step size adjustment strategy is used during iteration to accelerate convergence.

[0096] In the local execution control strategy, after the target power value is received, it first undergoes instruction validity verification to eliminate abnormal instructions that exceed the equipment tolerance. Local electrical quantity measurements employ sliding window filtering to eliminate instantaneous noise, and the actual output power is calculated based on instantaneous voltage and current sampling values ​​using a recursive least squares method. The first-order hysteresis filter time constant is dynamically adjusted according to the equipment's thermal inertia, and the proportional-integral controller parameters are tuned based on the system impedance characteristics. After the power compensation is generated, it undergoes output limiting processing to prevent over-adjustment from causing secondary oscillations, and finally, the executed instruction is verified using CRC to ensure transmission integrity.

[0097] The reinforcement learning optimization mechanism incorporates a reward signal design that considers three dimensions: voltage stability, frequency regulation effectiveness, and economic efficiency. A weighted fusion is used to generate a comprehensive reward value. The state vector is composed of run length distribution characteristics, run strength characteristics, and system voltage and current phasors. The action vector is defined as the adjustment amount of the threshold parameters in the mapping rule base and the weight coefficients of the objective function. Policy function training employs a combination of offline pre-training and online fine-tuning. In the offline phase, historical running data is used to initialize network parameters, while in the online phase, an ε-greedy strategy is used to balance exploration and utilization.

[0098] During the training of the deep Q-network, the experience replay buffer employs a priority sampling mechanism, giving higher sampling probabilities to samples with high temporal difference errors. Target value calculation uses a double-Q learning structure to avoid overestimation, with the discount factor set based on the system's inertial time constant. Network parameter updates utilize the Adam optimizer, with the learning rate dynamically decaying as training progresses. Target network parameter updates employ a soft update strategy, updating only a subset of parameters in each iteration to maintain training stability. Model pruning techniques are used during policy function deployment to reduce computational complexity and meet the timeliness requirements of real-time decision-making.

[0099] In its implementation, this invention utilizes synchronous measurement devices deployed at photovoltaic power plant grid connection points, electric vehicle charging stations, and key nodes in the distribution network to synchronously acquire three-phase voltage and current waveform data at a fixed sampling frequency. The measurement devices achieve time synchronization through a precision clock protocol, assigning a unified time stamp to each data point. The acquired data is then processed in real-time to generate a sequence of instantaneous photovoltaic output values, an electric vehicle charging power sequence, and a sequence of effective node voltage values, forming a multi-source synchronous measurement dataset with strict time alignment.

[0100] An adaptive noise-complete empirical mode decomposition (EMD) process is performed on photovoltaic power output sequences or node voltage sequences. This algorithm decomposes the non-stationary signal into several intrinsic mode function (IMF) components and residual terms by repeatedly injecting adaptive white noise and calculating the mean envelope. For each IMF component, its time series mean is used as a threshold to identify sample intervals continuously above or below the threshold and label them as independent runs. The number of sample points in each run is counted as the run length, and the sum of the squares of the deviations of the sample amplitudes within the run from the threshold is calculated as the run intensity. Based on the statistical results of all runs within the sliding time window, a feature set including a run length distribution histogram, average run intensity, and run information entropy is generated.

[0101] After normalizing the run-length features extracted from different monitoring points, principal component analysis (PCA) is used for dimensionality reduction and fusion to generate high-dimensional feature vectors representing the dynamic behavior of each power grid node. A static topology graph with electrical nodes as vertices and line connections as edges is constructed based on the power grid wiring diagram. The weights of the edges between nodes are then dynamically adjusted based on real-time power flow calculation results. The feature vectors are assigned as dynamic node attributes to the corresponding vertices of the topology graph, forming a spatiotemporal graph model where node attributes are updated with run-length features and edge weights change with power flow.

[0102] After the spatiotemporal graph is input into the prediction model, the spatial correlation strength between nodes is calculated using a graph attention network. Neighbor node features are weighted and aggregated using attention coefficients to update the vertex feature representation. The updated feature sequences are then input into prediction branches at three time scales (long, medium, and short), with each branch employing a temporal convolutional network with different dilation coefficients to extract temporal features. Multi-scale feature fusion is achieved through a cross-attention mechanism, and the output layer generates probability distribution predictions for photovoltaic power output, load demand, and dispatchable power of electric vehicles at different future time scales.

[0103] The mapping from run characteristics to safety constraints is achieved through a pre-defined rule base. The rule base defines the correspondence between different combinations of run characteristics and system safety risks. When the run intensity of the high-frequency component exceeds a threshold, constraints on the energy storage system's response rate are automatically generated. The mapping process matches real-time run characteristics with the rule base, triggering corresponding constraint generation logic and dynamically adjusting parameters such as power regulation rate and reserve capacity.

[0104] In the constructed two-layer model predictive control architecture, the upper-layer optimizer runs on an hourly cycle, integrating long-term forecast results and stable operating condition constraints to generate a baseline power plan with optimal economic efficiency as the objective. The lower-layer optimizer executes on a minute-by-minute cycle, integrating short- and medium-term forecasts, real-time measurement data, and fluctuating operating condition constraints, and employs distributed bar optimization to handle forecast uncertainties. The optimization problem is decomposed into photovoltaic, charging pile, and load subproblems using the alternating direction multiplier method, which is solved in parallel, outputting the power command sequence for each execution unit.

[0105] After the command is issued, the execution end adopts a local adaptive control strategy. Based on the received power command, the charging pile controller combines real-time sampled voltage and current data, smooths the command step through a first-order hysteresis filter, and uses a proportional-integral controller to compensate for execution deviations. This combination of global optimization and local fine-tuning effectively mitigates control errors caused by command transmission delays.

[0106] During the closed-loop optimization phase of the system, the actual power-voltage curves are spatiotemporally aligned with historical run characteristics to quantitatively evaluate the control effect under different fluctuation modes. A policy network is trained using reinforcement learning algorithms, with run characteristics and system state as inputs, dynamically adjusting constraint mapping rules and optimizing model weight parameters. This continuous learning mechanism based on actual operating data enables the system to adapt to slow time-varying factors such as changes in grid topology and the increasing penetration rate of new energy sources.

[0107] In the run-length analysis stage, a sliding window is used to calculate the local mean to adapt to the characteristics of non-stationary signals, and a minimum run-length threshold is set to eliminate transient noise interference. The run-length information entropy is calculated based on the probability distribution of run-length occurrence, and the Shannon entropy formula is used to quantify the randomness of the sequence. During the spatiotemporal graph construction process, the initial value of the edge weights is set to the reciprocal of the line impedance, and the weight signs are dynamically adjusted according to the real-time power flow direction. In the cross-attention mechanism of the prediction model, the short-term branch feature vector is used as the query vector, and the long-term branch feature vector is used as the key-value pair, achieving feature interaction through dot product attention.

[0108] The constraint mapping rule base adopts a production rule representation, where the antecedent of a rule is a logical combination of run-length features, and the consequent is a constraint generation instruction. The matching process uses a fuzzy matching algorithm to handle boundary cases, and the constraints support various types, including inequality constraints, equality constraints, and logical constraints. The effective duration of dynamic safety constraints is bound to the time window of the run-length features, ensuring timely matching between the constraints and the fluctuation features.

[0109] The lower-level rolling optimization employs a split-blob bar optimization method to handle prediction uncertainties, constructing a fuzzy set centered on the prediction probability distribution. The worst-case performance indicators are selected as the maximum system voltage deviation and the extreme value of frequency fluctuation as optimization objectives. Subsystems decomposed by the alternating direction multiplier method are connected to global coordination variables via Lagrange multipliers, and the augmented Lagrange method is used to handle coupling constraints. Convergence is determined based on a dual criterion of the original residual and the dual residual, and an adaptive step-size adjustment strategy is adopted during the iteration process.

[0110] The target power value in the local control loop undergoes command validity verification, eliminating abnormal commands that exceed equipment tolerances. Actual output power is calculated using a recursive least squares method based on instantaneous voltage and current sampling values. The first-order hysteresis filter time constant is dynamically adjusted according to the equipment's thermal inertia, and the proportional-integral controller parameters are tuned based on the system impedance characteristics. After power compensation is generated, it undergoes output limiting processing, and the final executed command is verified through a verification mechanism to ensure transmission integrity.

[0111] The design of the reinforcement learning reward signal incorporates multiple dimensions, including voltage stability, frequency regulation effectiveness, and economic efficiency. The state vector is composed of run-length characteristics and system electrical quantities, while the action vector is defined as the adjustment amount of the mapping rule parameters and the weights of the optimized model. Policy training combines offline pre-training with online fine-tuning, and the experience replay buffer employs a priority sampling mechanism. Deep Q-network training utilizes a double-Q learning structure to avoid overestimating value, and network parameter updates employ an adaptive optimization algorithm. The target network parameters use a soft update strategy to maintain training stability. During policy function deployment, model pruning techniques are used to reduce computational complexity and meet the timeliness requirements of real-time decision-making.

Claims

1. A method for coordinated scheduling of optical vehicles and loads based on multi-timescale matching, characterized in that, include: Step 1: Perform data acquisition to obtain synchronous measurement data of photovoltaic power output, electric vehicle charging power, and grid node voltage; Step 2: Perform run analysis on the photovoltaic output data or grid node voltage data in the synchronous measurement data to extract run features, including run length distribution and run intensity; Step 3: Using the extracted run features, generate feature vectors describing the dynamic attributes of the power grid topology nodes. Combine the power grid topology connection relationship with real-time power exchange data to construct a spatiotemporal graph with dynamic node attributes and edge weights. Step 4: Input the spatiotemporal diagram into the prediction model and output the photovoltaic power output prediction, load demand prediction and electric vehicle dispatchable power prediction at different time scales in the future. Step 5: Map the extracted run features to dynamic safety constraints and construct a model predictive control rolling optimization model. The rolling optimization model takes the photovoltaic output prediction, load demand prediction, electric vehicle dispatchable power prediction and the dynamic safety constraints as inputs and solves the model in a rolling manner to generate power command sequences for photovoltaic inverters, electric vehicle charging piles and controllable loads. Step 6: Issue and execute the power command sequence; collect the actual power and voltage data of the system after executing the power command sequence; use the collected actual power and voltage data of the system and the run features extracted by the run analysis to adjust the mapping relationship between the run features and the dynamic safety constraints, and adjust the parameters of the rolling optimization model.

2. The optical-vehicle-load cooperative scheduling method based on multi-timescale matching according to claim 1, characterized in that, Step 2 includes: For the photovoltaic power output data or grid node voltage data in the synchronous measurement data, perform adaptive noise complete empirical mode decomposition to obtain multiple intrinsic mode function components; For each obtained intrinsic mode function component, the time series mean of the component is calculated, and the mean is used as a threshold. In the time series of the component, segments that are continuously above or continuously below the threshold are detected, and each segment is marked as a run; The number of data points included in each run is counted as the run length; the sum of squares of the deviations of the amplitudes of all data points in each run from the threshold is calculated as the run intensity. Based on the run length and run intensity of all runs within a time window, the run length distribution, average run intensity, and run information entropy are determined. The run characteristics include the run length distribution, average run intensity, and run information entropy.

3. The optical-vehicle-load cooperative scheduling method based on multi-timescale matching according to claim 2, characterized in that, Step 3 includes: The run features extracted from the synchronous measurement data are collected, and normalization and principal component analysis are performed on the collected run features to generate the feature vector of each monitoring node in the corresponding power grid topology. Obtain the topological connections of the power grid, and construct a static topology graph with electrical nodes as vertices and electrical connections as edges; Calculate the power flow values ​​between nodes based on the real-time power exchange data in the synchronous measurement data; The weights of the corresponding edges in the static topology graph are dynamically adjusted using the calculated power flow values. The generated feature vectors are assigned as node attributes to the corresponding vertices in the static topology graph to obtain the spatiotemporal graph.

4. The optical vehicle-load cooperative scheduling method based on multi-timescale matching according to claim 3, characterized in that, Step 4 includes: Based on the attributes of vertices and the weights of edges in the spatiotemporal graph, the attention coefficient between a vertex and its neighboring vertices is calculated. The attention coefficients are used to weight the attributes of neighboring vertices and update the feature representation of the vertices. The updated vertex feature sequence is input into three processing branches: long-term, medium-term, and short-term. In each branch, a temporal convolutional network is applied to process the feature sequence. The dilation coefficients of the temporal convolutional networks in the long-term, medium-term, and short-term branches are different. Cross-attention calculation is performed between the long-term processing branch and the short-term processing branch to fuse the features output by the long-term branch and the features output by the short-term branch. The merged features are decoded to output the probability distribution of photovoltaic power output, load demand, and dispatchable power of electric vehicles at different time scales in the future.

5. The optical-vehicle-load cooperative scheduling method based on multi-timescale matching according to claim 4, characterized in that, Step 5 includes: A pre-defined mapping rule base is provided, which defines the mapping relationship between run-length characteristics and system safe operation boundaries. The extracted run-length features are matched with the run-length features defined in the mapping rule base; When a match is successful, at least one constraint is generated based on the mapping relationship, and the dynamic security constraint includes the generated constraint.

6. The optical-vehicle-load cooperative scheduling method based on multi-timescale matching according to claim 5, characterized in that, Step 5 further includes: The upper-level optimization calculation is performed in the first cycle. The input of the upper-level optimization calculation includes the long-term scale portion of the photovoltaic output forecast, load demand forecast and electric vehicle dispatchable power forecast, as well as the dynamic security constraints triggered by the long run characteristics in the run length distribution. The upper-level optimization calculation aims at optimal economic efficiency and satisfies the physical constraints of the power grid, and outputs the baseline power plan for future periods. The lower-level rolling optimization calculation is performed in a second cycle, which is shorter than the first cycle. The inputs to the lower-level rolling optimization calculation include the short-to-medium-term scale portions of the photovoltaic output forecast, load demand forecast, and electric vehicle dispatchable power forecast, the real-time power data in the synchronous measurement data, the baseline power plan, and dynamic safety constraints triggered by the high run intensity characteristics in the run intensity. Within each second cycle, a real-time optimization problem is constructed. The objective of the real-time optimization problem is to minimize the deviation between the actual power and the baseline power plan and to suppress power fluctuations. The constraints include grid physical constraints and dynamic security constraints triggered by the high run-length strength characteristics of the current input. The real-time optimization problem is solved to obtain the power command sequence.

7. The optical-vehicle-load cooperative scheduling method based on multi-timescale matching according to claim 6, characterized in that, Solving the real-time optimization problem to obtain the power command sequence includes: To address the uncertainty parameters introduced by prediction in the real-time optimization problem, a split-Brow bar optimization method is used. Construct a fuzzy set, which is centered on the probability distribution of the corresponding short-to-medium-term scale in the photovoltaic power output prediction, load demand prediction and electric vehicle dispatchable power prediction; Based on the fuzzy set, the objective function of the real-time optimization problem is reconstructed as minimizing the system performance deviation and fluctuation in the worst case. By applying the alternating direction multiplier method, the reconstructed real-time optimization problem is decomposed into multiple subproblems, each subproblem corresponding to a type of adjustable resource; The multiple sub-problems are computed in parallel, and information is exchanged through global coordination variables during the computation process. The solutions to the sub-problems are iteratively updated until the solutions converge. The converged solutions are then used as the power command sequence.

8. The optical-vehicle-load cooperative scheduling method based on multi-timescale matching according to claim 7, characterized in that, Step 6 includes: Issue the power command sequence; At the command execution end, the power command sequence is received and the target power value is extracted. Acquire local electrical quantity measurements at a rate higher than the power command sequence update frequency; A first-order hysteresis filtering algorithm is used to smooth the target power value to obtain a smoothed power command. The actual output power is calculated based on the collected local electrical quantity measurements. Calculate the deviation between the actual output power and the smoothed power command, and apply a proportional-integral control algorithm to the deviation to generate a power compensation amount; The power compensation amount is superimposed on the smoothed power command to obtain the final executed power command.

9. The optical-vehicle-load cooperative scheduling method based on multi-timescale matching according to claim 8, characterized in that, Step 6 also includes: The collected actual system power and voltage data are aligned with the run-length features extracted during the execution of the power command sequence by timestamps; Based on the aligned data, the execution effect of the power command sequence under different run characteristics is evaluated, and a reward signal is generated; Define a state vector, which includes a run feature vector and key system state variables; define an action vector, which includes adjustments to parameters in the mapping rule base or adjustments to the weights of the objective function of the rolling optimization model. Using the generated reward signal, a policy function is trained through a reinforcement learning algorithm, with the training objective being to maximize the expected value of future cumulative rewards. Using the trained policy function, an action vector is output based on the current state vector. The parameters of the mapping rule base and the rolling optimization model are then adjusted based on the output action vector.

10. The optical-vehicle-load cooperative scheduling method based on multi-timescale matching according to claim 9, characterized in that, The policy function trained by the reinforcement learning algorithm includes: Initialize the parameters of the deep Q-network, which is used to evaluate the value of state-action pairs; Initialize the experience replay buffer; In each training iteration, the following operations are performed: a batch of experience tuples are sampled from the experience replay buffer. Each experience tuple includes a historical state vector, an action vector executed, a reward signal obtained, and a next state vector. Input the sampled next state vector into the current depth Q network to calculate the maximum predicted action value for the next state; Calculate the target value, which is the sum of the reward signal in the sampled empirical tuple and the discounted maximum predicted action value for the next state; Input the historical state vector and the executed action vector from the sampled empirical tuple into the current depth Q network to obtain the current action value prediction; The temporal difference error algorithm is used to calculate the loss between the target value and the current action value prediction. The parameters of the current deep Q-network are updated using the calculated loss through backpropagation algorithm. Periodically update the parameters of an independent target depth Q-network; After training, the current deep Q-network is used as the policy function to select the action vector that maximizes the value of the predicted action based on the input state vector.

Images