An intelligent regulation system and method based on demand response and power compensation analysis conforming to aggregation
Patent Information
- Application Number
- CN202611045093.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-14
- Publication Date
- 2026-09-15
Smart Images
Figure CN122763489A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system load regulation technology, specifically to an intelligent regulation system and method based on demand response and conformity aggregation for power compensation analysis. Background Technology
[0002] In power systems where the proportion of renewable energy generation continues to increase, operational demands such as supply-demand balance, load fluctuation regulation, peak shaving and valley filling, and renewable energy consumption are constantly rising. Relying solely on generation-side regulation is no longer sufficient to meet the requirements for flexible grid operation. Demand response, through price signals, compensation incentives, or control commands, guides users to adjust their electricity consumption behavior, transforming dispersed demand-side resources into flexible regulation capabilities that can participate in dispatch. It has become an important means of power load regulation.
[0003] In existing technologies, load aggregators typically manage dispersed power load resources such as industrial loads, commercial loads, residential loads, charging loads, and energy storage loads. They determine response tasks based on grid-side control requirements and then issue compensation incentive information and control plans to users or equipment. Users or equipment participate in the response based on compensation conditions, electricity demand, and operating status. The load aggregator then performs compensation settlement and dispatch evaluation based on the response results.
[0004] However, existing load aggregation and control methods still have certain shortcomings in practical applications. On the one hand, the adjustable capacity reported by users or the response capacity predicted based on historical data often fails to accurately reflect the actual achievable capacity. User participation willingness, equipment operating status, response duration, start-up delay, and load recovery after control can all lead to deviations between the actual response and the expected results. On the other hand, existing methods typically focus on baseline load estimation or single response prediction, failing to fully consider the uncertainty of the response state load under different compensation and control conditions, resulting in insufficient reliability of compensation and control schemes.
[0005] Meanwhile, existing compensation mechanisms mostly adopt fixed compensation, tiered compensation, or simple settlement methods based on response volume, lacking sufficient linkage between compensation parameters and load control parameters. In actual implementation, if resource allocation plans are determined solely based on predicted response capacity or historical average response results, problems such as over-allocation of some resources, insufficient response, high compensation costs, or insufficient probability of achieving control targets can easily occur, affecting the stability and economy of load aggregation control.
[0006] Therefore, how to more accurately assess the dispatchable response capability of each load resource under different compensation and control conditions in the scenario of power load aggregation, and combine response reliability, compensation constraints and control task requirements to form an executable aggregated load control scheme, has become a technical problem that urgently needs to be solved in the field of demand response and load aggregation control. Summary of the Invention
[0007] The purpose of this invention is to provide an intelligent control system and method for power compensation analysis based on demand response and conformity aggregation, which solves the problems in the existing power load aggregation control process, such as large deviations between the response capacity reported by users or predicted by models and the actual executable capacity, insufficient linkage between compensation parameters and control parameters, and difficulty in effectively participating in control decision-making regarding the reliability of load resource performance.
[0008] To achieve the above objectives, this invention provides an intelligent control method for power compensation analysis based on demand response and conformity aggregation, comprising:
[0009] Obtain load control data for multiple power load resources to be aggregated, and determine the current control task and compensation constraints;
[0010] Multiple sets of candidate compensation and control parameters are set for each power load resource. Based on the load control data, multiple sets of baseline load prediction data are generated under the same operating conditions. Response state load prediction data corresponding one-to-one with the baseline load prediction data is generated for each set of candidate compensation and control parameters.
[0011] Based on the difference between the corresponding baseline load forecast data and the response state load forecast data, the response capacity distribution of each power load resource under each set of candidate compensation and control parameters is formed, and the lower bound of the response capacity that satisfies the preset completion probability in each response capacity distribution is determined.
[0012] The reliability of each power load resource's performance is determined based on historical response events that match the current control task. The response capacity lower bound corresponding to each group of candidate compensation control parameters is corrected by combining response duration, start-up delay, and post-response recovery risk, so as to obtain the dispatchable response capacity corresponding to each group of candidate compensation control parameters.
[0013] Based on the current control tasks and compensation constraints, a target combination is selected from the candidate compensation control parameters and corresponding dispatchable response capacities of each power load resource, and the dispatchable response capacity in the target combination is used as a constraint to form and execute an aggregated load control scheme.
[0014] As a further improvement to the above technical solution, the load control data includes control demand data, historical load data of each power load resource, operating status data, and historical response event data; the current control task is determined based on the control direction, target control amount, response period, and preset completion probability in the control demand data, and the compensation constraint is determined based on the compensation budget or compensation range in the control demand data.
[0015] As a further improvement to the above technical solution, multiple sets of candidate compensation and control parameters are set for each power load resource, including: determining the adjustable boundary of each power load resource based on the current control task, compensation constraints and operating status data of each power load resource; forming multiple sets of candidate compensation and control parameters within the adjustable boundary, each set of candidate compensation and control parameters including a candidate compensation parameter and a corresponding candidate control parameter.
[0016] As a further improvement to the above technical solution, the steps for generating the baseline load forecast data and the response state load forecast data include: generating baseline load forecast data for the same forecast period without compensation control based on the historical load data and operating status data of each power load resource; associating the operating status data corresponding to the baseline load forecast data with each group of candidate compensation control parameters to generate corresponding response state load forecast data, so that each group of response state load forecast data corresponds one-to-one with the baseline load forecast data under the same operating conditions.
[0017] As a further improvement to the above technical solution, it also includes: inputting historical load data and operating status data of each power load resource into a shared time-series feature extraction network to obtain common time-series features; inputting the common time-series features into a baseline load prediction branch to output the baseline load prediction data; and fusing the common time-series features with each group of candidate compensation and control parameters and inputting the result into a response status load prediction branch to output the corresponding response status load prediction data.
[0018] As a further improvement to the above technical solution, the step of determining the lower bound of the response capacity includes: for any power load resource and any set of candidate compensation control parameters, according to the adjustment direction of the current control task, performing difference processing on the corresponding baseline load forecast data and response state load forecast data to obtain multiple response capacity samples; forming a response capacity distribution based on the response capacity samples, and determining the lower bound of the response capacity corresponding to the set of candidate compensation control parameters according to the preset completion probability.
[0019] As a further improvement to the above technical solution, the steps for determining the performance credibility of each power load resource include: extracting historical control task characteristics and historical response effect characteristics from the historical response events of each power load resource; determining the event matching weight based on the degree of matching between the historical control task characteristics and the current control task; determining the event performance evaluation value based on the historical response effect characteristics; and weighting the event performance evaluation value based on the event matching weight to determine the performance credibility of each power load resource.
[0020] As a further improvement to the above technical solution, the step of obtaining the schedulable response capacity includes: determining a performance correction coefficient based on the performance credibility; determining a continuous correction coefficient based on the matching relationship between the response duration and the target response duration of the current control task; determining a delay correction coefficient based on the start delay and the allowed start time of the current control task; and determining a recovery correction coefficient based on the recovery risk after the response. The corresponding lower bound of the response capacity is corrected using the performance correction coefficient, continuous correction coefficient, delay correction coefficient, and recovery correction coefficient to obtain the schedulable response capacity corresponding to each group of candidate compensation control parameters.
[0021] The selection step of the target combination includes: selecting a target combination that includes the target compensation control parameters and the target schedulable response capacity from each group of candidate compensation control parameters and their corresponding schedulable response capacities, based on the degree of matching between the schedulable response capacity and the current control task, and the expected compensation cost determined by the candidate compensation control parameters.
[0022] As a further improvement to the above technical solution, after executing the aggregated load control scheme, the method further includes: acquiring the actual response data of each power load resource; determining the response deviation based on the actual response data and the target adjustment information corresponding to the aggregated load control scheme; and classifying the response deviation into at least one of the following according to the generation stage: baseline prediction deviation, response state prediction deviation, performance execution deviation, and post-response recovery deviation; and updating at least one of the following according to the type of response deviation: the generation parameters of the baseline load prediction data, the generation parameters of the response state load prediction data, performance credibility, and post-response recovery risk.
[0023] This invention also provides an intelligent control system for power compensation analysis based on demand response and conformity aggregation, comprising:
[0024] The data acquisition unit is used to acquire load control data of multiple power load resources to be aggregated, and to determine the current control task and compensation constraints.
[0025] The prediction processing unit is used to set multiple sets of candidate compensation and control parameters for each power load resource, generate multiple sets of baseline load prediction data under the same operating conditions based on the load control data, and generate response state load prediction data corresponding one-to-one with the baseline load prediction data for each set of candidate compensation and control parameters.
[0026] The response capacity determination unit is used to form the response capacity distribution of each power load resource under each set of candidate compensation and control parameters based on the difference between the corresponding baseline load forecast data and the response state load forecast data, and to determine the lower bound of the response capacity that satisfies the preset completion probability in each response capacity distribution.
[0027] The capacity correction unit is used to determine the performance reliability of each power load resource based on historical response events that match the current control task, and, in combination with response duration, start-up delay and post-response recovery risk, correct the lower bound of the response capacity corresponding to each group of candidate compensation control parameters to obtain the dispatchable response capacity corresponding to each group of candidate compensation control parameters.
[0028] The control execution unit is used to select a target combination from the candidate compensation control parameters and corresponding dispatchable response capacity of each power load resource based on the current control task and compensation constraints, and to form and execute the aggregated load control scheme with the dispatchable response capacity in the target combination as a constraint.
[0029] The beneficial effects of the intelligent control system and method for power compensation analysis based on demand response and conformity aggregation provided by this invention are as follows:
[0030] First, the present invention generates baseline load forecast data and response state load forecast data under the same operating conditions, and forms a response capacity distribution based on the difference between the two. This can reduce the capacity assessment deviation caused by direct comparison of load forecast results under different operating conditions and improve the reliability of response capacity calculation.
[0031] Second, this invention determines the lower bound of the response capacity and the schedulable response capacity for different candidate compensation and control parameters, thereby establishing a correlation between the selection of compensation and control parameters and the actual executable response capability of load resources, and avoiding the determination of control schemes based solely on user-reported capacity or single-point predicted values.
[0032] Third, this invention uses performance reliability, response duration, startup delay, and post-response recovery risk to correct the lower bound of response capacity, which can introduce historical performance and response execution risk into the schedulable capacity assessment process, thereby improving the executability of aggregated load control schemes.
[0033] Fourth, based on the current control task and compensation constraints, the present invention selects a target combination from candidate compensation control parameters and corresponding schedulable response capacities, which can meet the control requirements while taking into account the compensation constraints, thereby improving the stability and economy of load aggregation compensation control. Attached Figure Description
[0034] Figure 1 This is a flowchart illustrating the operation of the intelligent control method for power compensation analysis based on demand response and conformity aggregation in this embodiment of the invention.
[0035] Figure 2 This is a schematic diagram of the intelligent control system structure for power compensation analysis based on demand response and conformity aggregation in an embodiment of the present invention;
[0036] Figure 3This is a schematic diagram of the shared temporal feature extraction and dual-branch load prediction network structure in an embodiment of the present invention;
[0037] Figure 4 This is a schematic diagram of the engineering deployment of an intelligent control system for power compensation analysis based on demand response and conformity aggregation in an embodiment of the present invention;
[0038] Figure 5 This is a comparison chart of baseline load forecast data and response state load forecast data in an embodiment of the present invention;
[0039] Figure 6 This is a comparison chart of the response capacity distribution corresponding to different candidate compensation control parameters in the embodiments of the present invention.
[0040] Figure 7 This is a comparison diagram of the response capacity before and after correction in an embodiment of the present invention;
[0041] Figure 8 This is a diagram showing the target combination optimization results in an embodiment of the present invention;
[0042] Figure 9 This is a comparison chart of the load curves before and after the implementation of the polymerization load control scheme in this embodiment of the invention;
[0043] Figure 10 This is a comparison chart of the response deviation distribution and the effect before and after feedback update in an embodiment of the present invention. Detailed Implementation
[0044] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. It should be understood that the described embodiments are only for illustrating the present invention and should not be construed as limiting the scope of protection of the present invention. Other implementation methods obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention.
[0045] This invention provides an intelligent control method for power compensation analysis based on demand response and conformity aggregation. The power load resources managed by the load aggregation platform can be interruptible or transferable loads from industrial production lines, air conditioning and cold storage equipment in commercial buildings, charging loads at charging stations, distributed energy storage, residential temperature-controlled loads, or other electricity users capable of changing their active power within a predetermined time period. The platform distinguishes each power load resource by a resource identifier and stores its historical load, operating status, historical response records, and available control interfaces. The following example uses an aggregated resource pool consisting of forty-eight power load resources for illustration, with a data sampling period of fifteen minutes, and the prediction time resolution consistent with the sampling period.
[0046] In this implementation, the computational tasks can be performed by a single load aggregation and control server, or they can be split between a central server and an edge gateway for collaborative execution. The central server is responsible for data collection, load forecasting, response capacity distribution calculation, performance reliability correction, and target combination selection; the edge gateway is responsible for protocol conversion, command caching, on-site status verification, and execution result feedback. For directly controllable devices, the edge gateway converts the aggregated load control scheme into device control commands; for resources requiring user confirmation, the edge terminal sends compensation conditions and target adjustment information to the user-side application and issues control commands upon receiving confirmation.
[0047] like Figure 1 As shown, the intelligent control method for power compensation analysis based on demand response and conformity aggregation in this embodiment includes five processing stages: load control data acquisition and task determination, candidate compensation control parameter setting and paired load prediction, response capacity distribution and response capacity lower bound determination, joint correction of performance credibility and execution risk, and target combination selection and execution of aggregated load control scheme. Intermediate results are transmitted through task numbers, resource numbers, candidate group numbers, and scenario numbers in each stage, enabling data objects formed in the previous stage to serve as input for the next stage, while maintaining the correspondence between the same resource, the same candidate group, and the same operating scenario.
[0048] S101: Obtain load control data for multiple power load resources to be aggregated, and determine the current control task and compensation constraints accordingly. The current control task should at least include the control direction, target capacity, response start time, response end time, and preset completion probability; the compensation constraints should at least include the total compensation budget or the allowable unit compensation range. The output of this step is a unified task object (Task) and data indexes associated with each resource. Subsequent processing will retrieve the corresponding data using the task number and resource number.
[0049] S102 sets multiple sets of candidate compensation and control parameters for each power load resource. Without changing operating conditions such as weather, day type, forecast period, and equipment availability, the platform generates baseline load forecast data for the same resource without applying compensation and control. Then, it applies each set of candidate compensation and control parameters to the same operating conditions, generating corresponding response state load forecast data. Each set of data includes a resource number, candidate group number, scenario number, and forecast time to ensure that subsequent difference calculations do not involve different operating conditions.
[0050] S103, for each resource, each candidate group, and each scenario, the corresponding baseline load forecast data and response status load forecast data are subtracted according to the adjustment direction to obtain a response capacity sample. The response capacity samples from multiple scenarios constitute the response capacity distribution corresponding to that candidate group; then, a lower bound for the response capacity is extracted from this distribution according to a preset completion probability. The lower bound for the response capacity represents the lowest response capacity that can be relied upon under the preset completion probability, rather than the maximum value or mean of the samples.
[0051] S104: Events matching the current control task are selected from historical response events. The fulfillment reliability of each resource is determined based on the task similarity and actual execution effect of the historical events. Subsequently, fulfillment reliability, response duration, initiation delay, and post-response recovery risk are converted into correction factors to correct the lower bound of the response capacity for each candidate group, resulting in the schedulable response capacity corresponding to the candidate compensation control parameters for each group. This capacity characterizes the upper limit of the capacity that can be incorporated into the control decision after considering execution reliability.
[0052] S105: Based on the current control task and compensation constraints, the candidate compensation control parameters and their schedulable response capacity for each resource are jointly compared, and a target combination is selected. The target combination may include one or more resources to be invoked, as well as the target compensation control parameters and target schedulable response capacity corresponding to each invoked resource. The platform uses the schedulable response capacity in the target combination as the upper limit of the resource adjustment amount, generates an aggregated load control scheme, and issues it for execution.
[0053] In the above execution process, the same operating conditions mean that the external variables and resource status used for pairwise prediction remain consistent, and the candidate compensation control parameters are the main difference inputs between the two types of predictions; the response capacity distribution refers to the set of response capacity samples formed by multiple pairwise prediction scenarios under the same candidate group for the same resource; the target combination refers to the combination of the candidate compensation control parameters and the corresponding schedulable response capacity selected after task matching and compensation constraint screening.
[0054] In one specific implementation of the above method, the load control data includes control demand data, historical load data of each power load resource, operating status data, and historical response event data. Control demand data is generated by the dispatch interface or the task management interface of the load aggregation platform; historical load data is periodically collected by smart meters, energy management systems, or edge gateways; operating status data is reported by field controllers, equipment management systems, or user-side terminals; and historical response event data is archived by the event management service after each control operation.
[0055] Regulation demand data is transmitted using JSON objects, with example fields including task_id, direction, target_kw, start_time, end_time, completion_probability, budget_total, and compensation_range. Historical load data is stored in a time-series database using a long table structure with timestamps, or can be imported in batches from CSV or Parquet files; each record includes at least resource_id, timestamp, active_power_kw, and quality_code. Operational status data can be transmitted using JSON or Protobuf messages, and in addition to resource identifiers and timestamps, may also include available_flag, state_code, control_mode, soc, temperature, upper_limit_kw, and lower_limit_kw. Historical response events are stored separately in event tables and curve tables. The event table stores task conditions and evaluation results, while the curve table stores the actual power sequences before, during, and after the event.
[0056] Table 1. Example fields and uses of load control data
[0057]
[0058] After data access, the data acquisition service first converts all timestamps to a unified time zone and then aligns them by 15-minute boundaries. For short-term missing data, linear interpolation of adjacent valid values is used. Records with more than four consecutive missing sampling points are marked as missing and removed from the current prediction scenario. Power spikes are double-checked using both the device's rated power and adjacent rate of change. Records exceeding 120% of the rated power or the consecutive rate of change threshold are not directly used for model training but are retained in the abnormal data table. The cleaned data is then indexed using resource_id and timestamp.
[0059] In this example, the current regulation task is a reduction task, with the task period from 17:00 to 20:00, a target regulation amount of 64kW, and a preset completion probability of 0.90. The compensation budget cap is 340 yuan, and the unit compensation range is 0.35 yuan / kWh to 0.85 yuan / kWh. The task management service writes the above fields into the Task object and generates sixteen fifteen-minute prediction points based on the task period. The first twelve prediction points belong to the formal response period, and the last four prediction points are used to observe the recovery behavior after the response.
[0060] If the control demand data comes from multiple superior systems, the task management service can merge them according to task priority, start time, and resource usage conflicts. For tasks in opposite directions within the same time period, the task with the higher safety level is retained first; for tasks in the same direction, the target adjustment amounts are added together and the more stringent completion probability is taken. Compensation constraints are stored separately and not incorporated into the prediction input; they are only invoked when candidate compensation control parameters are generated and target combinations are selected.
[0061] When setting multiple sets of candidate compensation and control parameters for various power load resources, the platform first determines the adjustable boundaries of the resource based on the current control task, compensation constraints, and resource operating status. The adjustable boundaries include the permissible adjustment direction, available power range, longest continuous response time, earliest start-up time, and latest recovery time. For energy storage or electric vehicles, charging and discharging power is also limited based on the state of charge and off-grid time; for temperature-controlled loads, control quantities are limited based on equipment start-up / shutdown status and permissible setting range; for industrial loads, interruptible or transferable capacity is limited based on production plans and minimum continuous operating time.
[0062] In a typical implementation, the platform does not directly perform a full-space search for compensation prices and control quantities. Instead, it generates a finite candidate set within the adjustable boundaries. Candidate compensation parameters can be represented by unit compensation prices, fixed participation subsidies, or tiered coefficients, while candidate control parameters can be represented by adjustment ratios, control setpoints, response start offsets, and durations. The system generates candidate compensation parameters from low to high compensation intensity and simultaneously configures corresponding control parameters based on the historical sensitivity of resources to compensation signals, ensuring that each candidate set is a complete parameter combination that can be actually deployed.
[0063] Table 2 Examples of candidate compensation and control parameters for a certain power load resource
[0064]
[0065] Each candidate group is encoded as a fixed-length vector. For example, a candidate group vector can be written as [c_rate, u_ratio, delay_set, duration_set, direction_code], where continuous quantities are normalized according to their respective upper and lower limits, and the adjustment direction is discretely encoded. The candidate group number, parameter vector, and resource number together constitute the candidate_key. Subsequent response state prediction data, response capacity samples, capacity lower bound, and schedulable response capacity are all associated with the candidate_key to avoid parameter and capacity mismatch when selecting target combinations.
[0066] When the resource status changes after the candidate group is generated, the edge gateway can set a status version number. If the status version before the prediction starts is inconsistent with the status version before the instruction is issued, the control execution unit does not directly use the original candidate group, but re-verifies the adjustable boundary; if the change is still within the allowable range of the candidate group, the control parameters are corrected and execution continues; if it exceeds the adjustable boundary, the resource is removed from the target combination and a backup resource is called.
[0067] Based on the aforementioned candidate groups, the platform generates baseline load forecast data and response state load forecast data for each resource within the same forecast period. The baseline load forecast represents the load trajectory under the current weather, day type, historical load, and operating status conditions, without any candidate compensation control parameters applied. The response state load forecast represents the load trajectory formed after applying a certain candidate group under the same external conditions and initial operating status. The key to generating these pairs is to use the same scenario number for external random disturbances, ensuring that the difference between the two curves primarily reflects the impact of the candidate compensation control parameters.
[0068] In practice, the scenario generation service generates multiple scenarios based on weather forecast errors, fluctuations in user behavior, fluctuations in device availability, and historical model residuals. Each scenario contains a set of sixteen external variable sequences and a set of initial resource states. For candidate group Gj, the system first sets the candidate parameters to zero under scenario s to generate a baseline load prediction vector. Subsequently, keeping the external variables and initial state of scenario s unchanged, the parameter vector of Gj is loaded to generate the response state load prediction vector. Although different candidate groups can share the same baseline prediction curve, the database still maintains a reference to the baseline curve for each candidate group to maintain a one-to-one data relationship.
[0069] The model input tensor is constructed independently for each resource. The historical window uses ninety-six 15-minute sampling points from the past. The basic input includes active power, power change rate, time sine / cosine encoding, day of the week type, temperature, or other available operating status. For resources lacking a certain equipment status, the corresponding field is set to zero, and a missing value mask is provided. The predicted output is a power vector of sixteen sampling points, in kW. The output record includes at least resource_id, candidate_id, scenario_id, forecast_time, baseline_kw, and response_kw.
[0070] Before the generated results are written to the prediction results table, three checks are performed: First, the timestamps of the baseline load prediction data and the response state load prediction data must be completely consistent; second, the predicted power must not exceed the rated power of the resource and the state boundary; third, when the candidate control parameter is zero, the difference between the outputs of the two branches should be within the model tolerance range. Data that fails the checks will not be included in the response capacity distribution calculation and will trigger the model service to re-infer or use the most recent valid prediction result.
[0071] Baseline load forecast data and response state load forecast data can be generated by a shared time-series feature extraction network and two forecast branches. For example... Figure 3 As shown, the historical load sequence and operating status sequence are first processed for missing values, time alignment, and normalization, and then input into a shared temporal feature extraction network. The shared network consists of a residual temporal convolutional layer, a bidirectional temporal memory layer, and a multi-head attention layer; its output common temporal features are fed into the baseline load prediction branch and the response state load prediction branch, respectively. In addition to the common temporal features, the response state load prediction branch also receives the embedding vector of candidate compensation and control parameters, so the same common temporal feature can quickly generate response state load prediction results corresponding to multiple candidate groups.
[0072] Table 3 Example configuration of a dual-branch load forecasting network
[0073]
[0074] Before entering the network, the m-th continuous feature is standardized based on training set statistics. The standardized input preserves the relative variations of different resources while avoiding gradient bias caused by differences in the dimensions of power, temperature, and compensation parameters. Its common expression is:
[0075]
[0076] in, Let m be the m-th original feature of resource i at time t. and These are the mean and standard deviation of the feature in the training set, respectively. These are the standardized features. When If the value is less than the preset minimum value, the feature is treated as a constant feature and set to zero to avoid numerical amplification.
[0077] The residual temporal convolutional layer employs one-dimensional causal dilated convolution. The convolutional kernel only reads the current time step and its historical position, ignoring data after the prediction time step; the dilation coefficients of the three layers are set to 1, 2, and 4 respectively, allowing a smaller number of layers to cover a longer historical range. The convolutional output of the l-th layer at time t can be calculated using the following formula:
[0078]
[0079] in, For the first Layer at time The convolution output, The kernel length is 1. For the first The coefficient of thermal expansion of the layer, For the first Layer Each convolution weight, For bias, For activation function, This is the input feature corresponding to the historical position of the previous layer. The convolution output is added to the input of this layer through a residual connection; if the number of channels is inconsistent, a 1×1 convolution is first used to transform the number of channels. This structure is used to identify local patterns such as load ramp-up, peak precursors, periodic fluctuations, and short-term equipment start-up and shutdown.
[0080] The convolutional features are then fed into a bidirectional temporal memory layer. The forward memory unit reads the convolutional features sequentially, while the backward memory unit reads the same historical window in the opposite direction. The hidden states of both are concatenated to form the temporal representation of each historical moment. This bidirectional computation only occurs within the observed historical window and does not use future real data for the prediction period, thus preventing information leakage.
[0081] The multi-head attention layer generates a query vector based on the prediction task and calculates weights for features at each time step within the historical window. A common representation of a single attention head is:
[0082]
[0083] in, , and These are the query matrix, key matrix, and value matrix, respectively. The key vector dimension is used. Four attention heads learn intraday cycles, short-term changes, operational status, and regulatory patterns, respectively. These are concatenated and passed through a linear layer to obtain resources. In the scene 128-dimensional common temporal features .
[0084] The baseline load forecasting branch uses only common time-series features, while the response state load forecasting branch embeds candidate compensation and control parameters into a vector. Concatenate with common temporal features. The two branches can be represented as follows:
[0085]
[0086] in, Indicates the baseline load forecast branch, This indicates the response state load forecasting branch. Representing resources In the scene Common temporal characteristics under, Representing resources candidate groups Corresponding parameter embedding, and These represent the baseline load prediction vector and the response state load prediction vector, generated in pairs, respectively. The two branches share the temporal feature extraction network but have independent output layers, which maintains feature consistency under the same operating conditions while allowing candidate compensation and control parameters to change the response state prediction results.
[0087] The network training employs a joint loss. For each training sample, the baseline prediction error and the response state prediction error are calculated simultaneously, and the two branches are balanced using a weight λ. Scenarios with missing prediction labels are excluded using a mask. The joint loss can be written as:
[0088]
[0089] in, For the first The baseline true load vector of each sample. For the first The true load vector of response state corresponding to each historical response event. and The outputs of the two prediction branches are respectively. To ensure an effective number of training samples, The weights for predicting the loss in response state, This represents the L2 norm. The example uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 128, and a maximum training epoch of 150. The training stops early when the validation set loss no longer decreases for twelve consecutive epochs, and the model parameters with the lowest validation loss are saved.
[0090] Training data is divided into training, validation, and test sets in chronological order to avoid data leakage caused by random shuffling between adjacent time periods. Exemplary data includes 15-minute intervals of data from 48 resources over 60 consecutive days, with the first 42 days used for training, the following 9 days for validation, and the last 9 days for testing. Samples of already occurred response events are only used for training the response state branch, while samples from normal operating days are used for training the baseline branch. After the model is deployed, it is in ONNX or TorchScript format. The central server can perform batch inference, and edge computing terminals can load the quantized model for low-latency inference.
[0091] Table 4 Comparison Results of Exemplary Prediction Models
[0092]
[0093] The results in Table 4 illustrate a feasible engineering configuration: a shared network simultaneously learns unregulated states and candidate regulated states on the same historical features, which can reduce the scenario bias caused by two independent models. The above values are simulation results obtained according to the example data and parameter configuration described in this embodiment. In actual deployment, the number of network layers, hidden dimensions, and training rounds can be adjusted according to resource type, historical sample size, and prediction time domain.
[0094] To enable the aforementioned dual-branch prediction, capacity assessment, and compensation control decisions to operate continuously in the actual field, this implementation further adopts an engineering deployment structure that coordinates the central server and the edge access layer.
[0095] Figure 4 This illustrates one engineering deployment method for the system. The load aggregation and control center can be deployed on a cloud server, data center server, or local server within the campus. Internally, the center includes interface and permission services, a time-series database and event library, a load forecasting model service, a capacity distribution and correction service, a target combination optimization service, a message bus, and a monitoring and auditing service. Each service can run as an independent container or within the same application process; services exchange data objects associated with resource_id, candidate_id, and task_id through internal REST interfaces or message queues.
[0096] The sample configuration of the central server includes an eight-core processor, 32GB of memory, and an optional GPU accelerator card. The operating system is a common Linux distribution. Time-series data can be stored in TimescaleDB or other time-series databases, short-term task states can be stored in Redis, and the model service runs through a Python inference framework or a C++ inference engine. The above configuration is not a strict requirement; with limited resources, a single ordinary industrial computer can simultaneously perform data acquisition, forecasting, capacity calculation, and optimization.
[0097] The edge access layer includes edge gateways and user-side terminals. Edge gateways can employ quad-core ARM processors, 4GB to 8GB of memory, and local non-volatile storage, running protocol conversion, data caching, command integrity verification, timeout handling, and network outage resumption procedures. Communication between the center and the edge can be via HTTPS, MQTT over TLS, private networks, or virtual private networks; communication between the edge and field devices can use Modbus TCP / RTU, IEC 61850, DL / T 645, or open interfaces from the device manufacturer.
[0098] The resource layer may include industrial or commercial energy management systems, temperature-controlled load controllers, charging pile controllers, energy storage EMS / PCS, smart meters, and home gateways. Each resource control interface must support reading the current power and availability status, and support receiving at least one type of control information, such as target power, start / stop commands, setpoints, or response periods. For resources that do not support direct control, the system only issues target adjustment information and compensation conditions, and the user or local controller must complete the response independently.
[0099] like Figure 4 As shown, the operational status and actual response data are uploaded from the resource layer to the message bus via the edge access layer, and then written to the time-series database. The aggregated load control scheme is generated by the target combination optimization service and distributed via the message bus and encrypted communication network. The monitoring and auditing service records task creation, model version, candidate group, target combination, instruction signature, execution confirmation, and feedback update time, so that any response capacity and compensation decision can be traced back to the corresponding input data and model version.
[0100] The data acquisition unit can be implemented using interface and permission services, a message bus, and a database connection program; the prediction processing unit can be implemented using a candidate parameter generation program and a load prediction model service; the response capacity determination unit can be implemented using a scenario calculation program and a distribution statistics program; the capacity correction unit can be implemented using a historical event matching program, a credibility calculation program, and a risk correction program; and the control execution unit can be implemented using a target combination optimization service, an instruction generation program, and an edge gateway execution program. Each unit can share processors and memory, or it can be distributed across different computing nodes at the center and edge.
[0101] like Figure 2 As shown, the data acquisition unit transmits load control data and task objects to the prediction processing unit; the prediction processing unit outputs paired baseline load prediction data and response status load prediction data; the response capacity determination unit uses this to form the response capacity distribution and response capacity lower bound; the capacity correction unit combines historical response events and execution risks to obtain the schedulable response capacity; the control execution unit reads the schedulable response capacity corresponding to each candidate group and completes the target combination selection, scheme generation, and instruction issuance. All units are connected through a unified data identifier and can be deployed on the same server or separately on a central server and an edge gateway.
[0102] The lower bound of the response capacity is determined by taking into account the corresponding baseline load forecast data and the response state load forecast data. Figure 5A pairwise prediction example for resources under candidate group G3 is given. The solid line in the figure represents the baseline load prediction data without compensation control, and the dashed line represents the response state load prediction data obtained after applying G3. The task response period is from 17:00 to 20:00, and the difference between the two curves during this period forms the predicted response capacity for this scenario. The average difference during the response period is approximately 20.5 kW, and the maximum difference at any given time is approximately 22.0 kW. The small recovery increment that appears after 20:00 is used separately for post-response recovery risk calculation and is not included in the formal response capacity.
[0103] For downward adjustment tasks, the response capacity of a single scenario is the average effective power obtained by subtracting the response state load forecast from the baseline load forecast value within the response period; for upward adjustment tasks, the difference is calculated in reverse. If the response power is needed as the task quantity, the difference can also be multiplied by the sampling interval and then summed. This example uses the average power capacity, expressed as:
[0104]
[0105] in, For resources In the candidate group and scene The following response capacity sample, The set of sampling points within the response period. For set The number of sampling points in the middle This is the task direction coefficient, which is 1 for downward tasks and -1 for upward tasks. and They are time points The baseline load forecast and the response status load forecast are used. To avoid abnormal distribution caused by a reverse response, a negative difference can be marked as an unexecutable sample or truncated to zero, depending on the task rules.
[0106] This example generates 500 scenarios for each resource and each candidate group. Weather forecast errors are sampled based on historical residual distributions, device availability is sampled using Bernoulli samples based on recent online rates, and user behavior perturbations are sampled based on historical load residuals for the same time period. Each scenario uses a fixed random seed to form a reproducible set of scenarios, and the same perturbation is used for the same scenario number in both the baseline branch and the response status branch.
[0107] The response capacity distribution can be directly represented by an empirical distribution. Suppose that for the same resource and candidate group, S sample responses are obtained, and their empirical distribution function is:
[0108]
[0109] in, For resources In the candidate group The empirical distribution function of response capacity under the given conditions, The number of valid scenario samples. For indicator functions, For the first Response capacity samples for each scenario This represents the response capacity value to be evaluated. The preset completion probability is... At that time, the lower bound of the response capacity It should meet at least The proportion of the sample is not lower than this value, therefore it can be determined by the empirical distribution. Quantiles were obtained:
[0110]
[0111] Figure 6 The distribution of response capacity for the same resource under three candidate compensation and control parameters is shown. The mean response capacity of candidate group 1, candidate group 2, and candidate group 3 increases sequentially, indicating that the predictive response capability improves after the compensation and control strength is increased; however, the distribution width also changes, indicating that the execution uncertainty cannot be reflected solely by the mean. When the preset completion probability is 90%, the lower bounds of the response capacity of the three candidate groups are approximately 34.2kW, 48.2kW, and 56.9kW, respectively.
[0112] Table 5 Statistical results of response capacity of different candidate groups
[0113]
[0114] When the number of valid scenario samples is less than the preset minimum, the system does not output a high-confidence capacity lower bound, but instead expands the scenario sampling or adopts a conservative boundary. If the response capacity distribution shows obvious multi-peaks, it can be calculated in layers according to device availability or user acceptance status; if the proportion of unexecutable samples in the candidate group exceeds... If the response capacity of the candidate group is zeroed, it is removed from the target combination candidate set.
[0115] The reliability of each power load resource's performance is calculated based on historical response events that match the current control task. Historical control task characteristics may include control direction, target capacity as a percentage of the resource's rated power, response period, target duration, compensation level, and pre-event operating status. Historical response effect characteristics may include the ratio of actual response quantity to target quantity, start-up delay, compliance duration percentage, number of execution interruptions, and post-response recovery increment. After each historical event, the event management service extracts these characteristics from the actual power curve and writes them into the event characteristic table.
[0116] Table 6 Examples of Historical Response Event Characteristics
[0117]
[0118] The event matching program first converts the task features of the current control task and historical events into the same feature vector, and then calculates the weighted distance. Discrete features use a mismatch value of 0 or 1, continuous features use a normalized difference, and time-period features use the response window overlap rate. The distance between the k-th historical event and the current task can be expressed as:
[0119]
[0120] in, For resources The The weighted distance between historical events and current tasks The number of task features. For the current task Task characteristics For resources The The first historical event corresponds to the Task characteristics Assign weights to the corresponding features. Events with inconsistent directions can be directly excluded, or their distance can be set to a larger value. Event matching weights are formed based on distance, event freshness, and data quality:
[0121]
[0122] in, For resources The The event matching weight corresponding to each historical event and The weighted distance between historical events and the current task. and The time interval between the event and the current task. The time decay coefficient, This is the filtered set of historical events. The matching weights are normalized to a sum of 1, so that similar and more recent historical events contribute more to the credibility of current performance.
[0123] Historical response performance characteristics are converted into event fulfillment evaluation values using scoring rules. In the example, the actual response volume achievement rate, timely initiation rate, continuous compliance rate, and recovery safety rate are weighted and combined, with the results limited to a range of 0 to 1.
[0124]
[0125] in, For resources The The performance evaluation value of a historical event. Scoring is given based on the achievement of the actual response amount and the target adjustment amount. To score the startup delay, To achieve the target, a continuous percentage score is required. The response recovery score is calculated; in the formula, 0.40, 0.20, 0.25, and 0.15 are example weights for each scoring item, and the sum of the weights is 1. The performance credibility is obtained by weighting the event performance evaluation value with the event matching weights:
[0126]
[0127] in, For resources The credibility of performance. Match weights to events, This is the performance evaluation value for the event.
[0128] For example, resource R1 has twelve similar historical events, resulting in a weighted performance reliability of 0.88; resource R4 has a high achievement rate and low startup delay, with a performance reliability of 0.92; resource R5 recently experienced two startup timeouts and one significant recovery rebound, resulting in a performance reliability of 0.72. For newly added resources without sufficient historical events, a conservative prior value of 0.65 can be used, similar to that of other resources, and gradually updated after each actual execution; once the number of historical events reaches a preset threshold, the performance reliability can be switched to an individualized level.
[0129] After obtaining the performance credibility, the system determines correction coefficients based on performance credibility, response duration, startup delay, and post-response recovery risk. The performance correction coefficient monotonically increases with ρ_i; the sustainability correction coefficient is determined based on the ratio of the resource's sustainable response time to the task's target response time; the delay correction coefficient is determined based on the predicted startup delay and the task's allowed startup time; and the recovery correction coefficient is determined based on the post-event recovery peak and recovery energy. All coefficients are limited to the range of 0 to 1.
[0130] A commonly used joint correction method is to employ a weighted geometric combination, which preserves the independent impact of each risk factor while avoiding excessive reduction caused by directly multiplying multiple coefficients. The combined correction coefficient and schedulable response capacity can be expressed as:
[0131]
[0132]
[0133] in, For resources In the candidate group The overall correction factor is as follows: For the corresponding schedulable response capacity, In response to the lower bound of capacity, This is the performance correction factor. For continuous correction factors, This is the delay correction factor. To restore the correction coefficient; the sum of all indices is 1. The weights can be adjusted according to the task type; for example, for urgent peak shaving tasks, the weight of startup delay is increased, and for long-term adjustment tasks, the weight of duration is increased.
[0134] Figure 7 The calibration results for six resources under a set of candidate parameters are presented. The lower bounds of the response capacity for R1 to R6 are 52kW, 44kW, 36kW, 61kW, 28kW, and 48kW, respectively. The schedulable response capacities obtained after joint calibration are 43kW, 37kW, 29kW, 55kW, 20kW, and 41kW, respectively. R5, with lower fulfillment reliability, experiences the largest reduction, while R4, with higher fulfillment reliability and lower recovery risk, retains the highest proportion of capacity. This figure illustrates that the calibration results are not equivalent to a uniform reduction ratio, but rather vary with the historical performance of the resources and the task adaptation status.
[0135] The selection of the target combination is based on the schedulable response capacity and expected compensation cost corresponding to each candidate group. For a single resource, the platform compares the capacity matching degree, start-up time, and compensation cost of the candidate groups for the current task; for multiple resources, the platform must also meet constraints such as the aggregated capacity reaching the target adjustment amount, the total compensation not exceeding the budget, each resource selecting at most one candidate group, and resource availability. Mixed-integer programming, particle swarm optimization, genetic algorithms, or non-dominated sorting genetic algorithms can be used to solve this problem.
[0136] The example uses the non-dominated sorting genetic algorithm NSGA-II. Chromosomes use resources as genes, with each gene having a value of 0 indicating no use and values from 1 to J indicating selection of a candidate group for that resource. The initial population prioritizes generating individuals that meet the budget and resource availability requirements; during crossover, candidate groups are exchanged according to resources, and during mutation, values are re-selected from the effective candidate groups for that resource; each round calculates the task capacity gap and the expected compensation cost, and retains individuals based on the non-dominated level and crowding distance. The two main objectives can be represented as:
[0137]
[0138]
[0139] in, For the normalized task capacity deviation target, To normalize the expected compensation cost target, For resources Select candidate group binary variables, Adjustment amount for the current task objective. For resources In the candidate group The schedulable response capacity is below. To estimate compensation costs, To compensate for the budget cap, resource availability, candidate group uniqueness, and budget limits are treated as constraints; when task capacity allows for a certain degree of overcapacity, upper and lower tolerance ranges can be set.
[0140] Figure 8 Example results of target combination optimization are given. The horizontal axis represents the schedulable response capacity, the vertical axis represents the expected compensation cost, the scatter plot depth indicates the degree of matching with the current control task, and the broken line represents the non-dominated frontier. The target capacity of the example task is approximately 64kW. The target combination marked with a star corresponds to a schedulable response capacity of 64.2kW, an expected compensation cost of 312 yuan, and a task matching degree of 0.95. This satisfies the budget constraint of 340 yuan and avoids selecting combinations with lower costs but insufficient capacity or higher capacity but significantly increased compensation.
[0141] Once the target combination is formed, the control execution unit generates a machine-readable control plan. The plan message can be in JSON or Protobuf format and includes at least the following parameters: task_id, resource_id, candidate_id, direction, start_time, end_time, target_kw, compensation_parameter, control_parameter, expiration_time, and signature. Upon receiving the message, the edge gateway verifies the resource status, message signature, and validity period, limits target_kw to the schedulable response capacity corresponding to the target combination, and then converts it into a control command recognizable by the device.
[0142] Once the aggregated load control scheme is implemented, the edge gateway collects the actual power and execution status of each resource at intervals of one minute or less, and generates actual response data with a 15-minute prediction granularity. The central server simultaneously retains both the original high-frequency data and the aggregated data: the high-frequency data is used to identify startup delays and short-term interruptions, while the aggregated data is used to compare with target adjustment information, baseline load prediction data, and response status load prediction data.
[0143] The actual response capacity can be calculated based on the difference between the baseline prediction and the actual power during execution. To reduce the impact of noise at individual sampling points, the system can first calculate the median power within each 15-minute interval, then calculate the difference according to the task direction, and finally average it over the entire response period.
[0144]
[0145] in, For resources The actual response capacity in this mission, For the set of sampling points during the response period, This represents the number of sampling points in the set. This is the task direction coefficient. The time corresponding to the target combination Baseline load forecast, For a moment The actual power collected.
[0146] Figure 9 The load curves before and after the implementation of the aggregated load control scheme are presented. Before control, the curve formed a significant peak during the evening peak, and after control, it decreased between 17:00 and 20:30, with a moderate increase in transferable load during the early morning low period. In the example, the peak load during the response period decreased from 625.2kW to 572.3kW, a decrease of 52.9kW, or approximately 8.5%; the standard deviation of the daily load decreased from 66.7kW to 58.4kW, and the load fluctuation decreased by approximately 12.4%. After the response ended, only a small recovery increment occurred, and no new peak was formed.
[0147] Response deviations are categorized based on the generation stage. If the actual power before the event begins has consistently deviated from the baseline load forecast data, the main deviation is marked as baseline forecast deviation. If the baseline forecast was accurate before the event, but the actual response trajectory deviates from the response status load forecast data of the selected candidate group, it is marked as response status forecast deviation. If resources fail to start on time, fail to meet targets consistently, or exit midway, it is marked as performance execution deviation. If the load recovery magnitude or recovery time exceeds a preset threshold after the response ends, it is marked as post-response recovery deviation. An event can have multiple deviation labels simultaneously, and a deviation contribution ratio is assigned to each label.
[0148] In one engineering implementation, the system first calculates the segmented errors before, during, and after the event, and then determines the source of the deviation based on the time location and status log. The total response deviation can be written as the segmented contribution:
[0149]
[0150] in, The total response deviation, Baseline prediction bias, In response to state prediction bias, For deviations in contract performance, To recover from deviations after a response. If the source of the deviation cannot be uniquely determined, weights can be assigned based on the error before and after the event, the device status log, and the time overlap ratio, instead of writing all errors to all updated objects simultaneously.
[0151] Baseline prediction bias is used to update the shared temporal feature extraction network and the baseline load prediction branch; response state prediction bias is used to update the candidate parameter embedding layer and the response state load prediction branch; performance execution bias is used to update the event performance evaluation value and performance credibility; and post-response recovery bias is used to update the recovery risk model. Model parameters can be fine-tuned offline after accumulating a certain number of new events, or updated online with a small learning rate. A common gradient update method is:
[0152]
[0153] in, and These are the model parameters before and after the update, respectively. To update the learning rate, This indicates that the gradient is calculated with respect to the model parameters. The loss function corresponds to the deviation type. Performance credibility can be updated using an exponential sliding method, meaning that the evaluation value of new events has a higher weight, while the old credibility retains a certain inertia; recovery risk can be updated based on the recovery peak, recovery energy, and duration of the most recent events. All updates record the data window, model version, and update time. The updated model is first validated on replay data before replacing the online model.
[0154] Figure 10 The distribution of absolute response bias rates before and after feedback updates is presented. Before the update, the mean bias rate was approximately 9.7%, and the 95th percentile was approximately 18.3%. After classifying the biases and updating the corresponding parameters separately, the mean bias rate decreased to approximately 5.4%, and the 95th percentile decreased to approximately 10.7%, with the distribution shrinking towards the low-bias region. This result indicates that targeted updates based on the bias generation stage can avoid parameter oscillations caused by updating the same model for all errors.
[0155] Table 7. Exemplary verification results of the complete control chain.
[0156]
[0157] Table 7 shows the comparison results of this implementation method on example data. When using only single-point prediction, the system tends to directly use the predicted mean as the available capacity, resulting in a high overcommitment rate. After introducing response capacity distribution and capacity lower bound, the task completion rate improves. After further adding feedback on fulfillment credibility, duration, start-up delay, recovery risk, and deviation classification, the actual response is more consistent with the target adjustment information. The above results are used to illustrate the feasibility of each processing step; the actual values will vary with resource composition, task scale, and compensation rules.
[0158] During online operation, model updates can be decoupled from task execution. The task execution process always uses the released stable model version, and feedback data enters the training area. The new version is switched to during non-task periods only after passing historical event replay, capacity coverage verification, and budget constraint verification. If the new version predicts abnormally or communication is interrupted, the system reverts to the previous stable version and uses the most recent effective schedulable response capacity or a preset conservative capacity to maintain basic controllability.
[0159] Without altering the aforementioned processing relationships, scenario generation can be achieved through Monte Carlo sampling, Latin hypercube sampling, or historical residual resampling; the shared temporal feature extraction network can employ temporal convolutional networks, recurrent neural networks, or combinations thereof; and the objective combination can be solved using deterministic programming, heuristic algorithms, or multi-objective optimization algorithms. Each replacement method must maintain a one-to-one correspondence between baseline load forecast data and response state load forecast data under the same operating conditions, and maintain the correlation between schedulable response capacity and candidate compensation and control parameters.
[0160] Before the aggregated load control scheme is issued, the central server performs a task-level consistency check. The check includes the task identifier, resource identifier, candidate group identifier, adjustment direction, execution time period, target capacity, compensation parameters, control parameters, message version, and validity period. The system verifies one by one whether the resources in the target combination are still available for use, whether the corresponding candidate group exists, whether the target capacity does not exceed the schedulable response capacity of the candidate group, whether the execution time periods of each resource overlap with the current task, and whether the cumulative estimated compensation does not exceed the compensation constraint. If the same resource is selected repeatedly in the same time period, only the scheme record with the newest version number and valid signature is retained; other records are placed in a conflict queue awaiting manual review or automatic re-optimization.
[0161] Task-level capacity verification employs a two-level tolerance. The first level is single-resource tolerance, used to limit the proportion of the target adjustment amount relative to the schedulable response capacity in the target combination. The second level is aggregation tolerance, used to limit the deviation between the sum of the target adjustment amounts of all invoked resources and the target adjustment amount of the current control task. For example, the single-resource target adjustment amount can be set to 0.85 to 1.00 times its target schedulable response capacity, and the aggregated capacity is allowed to be between 0.98 and 1.05 times the task target. If the capacity is insufficient, the optimizer prioritizes supplementing from the valid candidate group of uninvoked resources; if the capacity is excessive, the target amount of high-cost resources is reduced sequentially according to the joint ranking of expected compensation cost, recovery risk, and matching degree.
[0162] The edge gateway employs an idempotent execution mechanism to handle duplicate commands. Each control message is uniquely keyed by task_id, resource_id, and command_seq. The gateway stores the most recently executed sequence number and execution status in a local lightweight database. Messages with the same or older sequence numbers are not repeatedly written to the device control register; device-side commands are only generated after a new sequence number passes signature, timestamp, and parameter boundary verification. For loads that can be reduced, target power can be converted into start-stop combinations, power limits, or temperature setpoints; for loads that can be transferred, it can be converted into start-up time and runtime; for charging and discharging equipment, it can be converted into active power commands and state-of-charge protection boundaries. These conversions are performed by the resource adapter, while the central side still manages them using unified target adjustment amounts and candidate control parameters.
[0163] In the event of a communication failure, the gateway performs tiered degradation based on the remaining task duration and local status. If the link is briefly interrupted while the task instruction is still valid, the gateway continues to execute the most recently confirmed instruction and caches high-frequency measurement data locally; once the link is restored, the data is retransmitted in chronological order. If the link interruption exceeds a preset duration, the gateway no longer increases the adjustment range but maintains the resources at the most recent safe operating point or exits control at a preset slope. If the device status, power boundary, or safety protection conditions change, the gateway immediately prioritizes safety constraints over control objectives, terminates the remaining instructions for the corresponding resources, and reports the unavailability status, termination time, and termination reason to the central server. Upon receiving the status, the central server recalculates the remaining capacity gap of the resources.
[0164] The model inference service and optimization service run in independent containers, and model files, feature configurations, normalization parameters, and candidate parameter encoding tables are bound by a unified version list. When loading a model, the online service verifies the model hash value, the number of input fields, the time window length, and the output dimension; if any of these are inconsistent, the model is rejected for release. Model switching uses a dual-instance approach: the new instance first receives mirror traffic and generates shadow prediction results. Only after several consecutive prediction cycles that the error, capacity coverage, and inference latency requirements are met does it replace the current instance. The switching process retains the previous stable version to allow for rapid rollback in case prediction results exceed physical boundaries, output is missing, or inference latency is abnormal.
[0165] Acceptance testing can be conducted from five aspects: data integrity, prediction accuracy, capacity reliability, optimization feasibility, and execution real-time performance. Data integrity testing verifies the missing rate and time alignment rate of load curves, status fields, and event records; prediction accuracy testing calculates the mean absolute error of baseline load prediction data and response status load prediction data respectively; capacity reliability testing measures the proportion of events where the actual response capacity reaches the lower bound of the response capacity and the target schedulable response capacity; optimization feasibility testing verifies whether budget, candidate group uniqueness, and resource availability constraints are met; and execution real-time performance testing records the latency of each stage: task reception, model inference, target combination solution, instruction issuance, and terminal confirmation. An exemplary acceptance threshold can be set as follows: key field integrity rate not less than 99%, prediction interface success rate not less than 99.5%, capacity lower bound experience coverage not less than the preset completion probability minus the allowable error, and single target combination solution time not exceeding the control preparation window.
[0166] In a stress test environment containing one thousand logical load resources, the central server generates scenarios in batches with a 15-minute granularity, while edge devices upload statuses with a 1-minute granularity. When the shared temporal feature extraction network performs inference on a batch of one thousand resources, it can perform parallel batching by resource type or gateway partitioning; for objective combination solving, unusable candidate groups are first eliminated, and then multi-objective optimization is performed on the remaining combinations. During the test, CPU usage, memory usage, message backlog, database write latency, and end-to-end task generation time are recorded. When the number of load resources increases, it can be expanded by adding inference worker processes and by task sharding, without changing the correspondence between the candidate compensation and control parameters of each resource and the schedulable response capacity.
[0167] The system saves a replayable audit package for each load control task. The audit package includes a snapshot of the original load control data, data cleaning logs, model version, candidate compensation control parameter set, scenario random seed, response capacity distribution summary, capacity lower bound, correction coefficient, target combination, issued message, equipment acknowledgments, and feedback update results. The audit package is stored in an append-only manner, with key fields having calculated summary values and associated with the task identifier. In case of anomalies or when verification is required, intermediate results can be regenerated in an isolated environment using the same inputs and version to verify the capacity lower bound, target combination, and actual execution deviation, ensuring that the entire control chain can pinpoint the specific data, model, and rule version.
[0168] For different resource types such as industrial loads, commercial buildings, residential loads, charging facilities, and energy storage devices, resource status resolvers and control adapters can be configured separately. However, the unified objects they output to the central server all include timestamps, actual power, availability status, adjustable direction, adjustable boundaries, startup status, and recovery status. Thus, physical differences on the device side are confined to the adaptation layer, while the pairwise load forecasting, response capacity distribution, performance reliability correction, and target combination selection on the central side remain consistent. When adding a new resource type, only field mapping, physical boundary verification, and instruction conversion rules need to be supplemented; the aforementioned core processing chain does not need to be changed.
[0169] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, or combinations made to the above embodiments within the spirit and principles of the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention shall be determined by the scope defined in the claims.
Claims
1. A smart control method for power compensation analysis based on demand response and conformity aggregation, characterized in that, include: Obtain load control data for multiple power load resources to be aggregated, and determine the current control task and compensation constraints; Multiple sets of candidate compensation and control parameters are set for each power load resource. Based on the load control data, multiple sets of baseline load prediction data are generated under the same operating conditions. Response state load prediction data corresponding one-to-one with the baseline load prediction data is generated for each set of candidate compensation and control parameters. Based on the difference between the corresponding baseline load forecast data and the response state load forecast data, the response capacity distribution of each power load resource under each set of candidate compensation and control parameters is formed, and the lower bound of the response capacity that satisfies the preset completion probability in each response capacity distribution is determined. The reliability of each power load resource's performance is determined based on historical response events that match the current control task. The response capacity lower bound corresponding to each group of candidate compensation control parameters is corrected by combining response duration, start-up delay, and post-response recovery risk, so as to obtain the dispatchable response capacity corresponding to each group of candidate compensation control parameters. Based on the current control tasks and compensation constraints, a target combination is selected from the candidate compensation control parameters and corresponding dispatchable response capacities of each power load resource, and the dispatchable response capacity in the target combination is used as a constraint to form and execute an aggregated load control scheme.
2. The intelligent control method for power compensation analysis based on demand response and conformity aggregation as described in claim 1, characterized in that, The load control data includes control demand data, historical load data of each power load resource, operating status data, and historical response event data; the current control task is determined based on the control direction, target control amount, response period, and preset completion probability in the control demand data; and the compensation constraint is determined based on the compensation budget or compensation range in the control demand data.
3. The intelligent control method for power compensation analysis based on demand response and conformity aggregation according to claim 2, characterized in that, Multiple sets of candidate compensation and control parameters are set for each power load resource, including: determining the adjustable boundary of each power load resource based on the current control task, compensation constraints and operating status data of each power load resource; forming multiple sets of candidate compensation and control parameters within the adjustable boundary, each set of candidate compensation and control parameters including a candidate compensation parameter and its corresponding candidate control parameter.
4. The intelligent control method for power compensation analysis based on demand response and conformity aggregation as described in claim 2, characterized in that, The steps for generating the baseline load forecast data and the response status load forecast data include: generating baseline load forecast data for the same forecast period without compensation control based on the historical load data and operating status data of each power load resource; associating the operating status data corresponding to the baseline load forecast data with each group of candidate compensation control parameters to generate corresponding response status load forecast data, so that each group of response status load forecast data corresponds one-to-one with the baseline load forecast data under the same operating conditions.
5. The intelligent control method for power compensation analysis based on demand response and conformity aggregation according to claim 4, characterized in that, Also includes: Historical load data and operating status data of each power load resource are input into a shared time-series feature extraction network to obtain common time-series features; The common time-series features are input into the baseline load forecasting branch, and the baseline load forecasting data is output. The common time-series features are fused with the candidate compensation and control parameters of each group and then input into the response state load prediction branch, which outputs the corresponding response state load prediction data.
6. The intelligent control method for power compensation analysis based on demand response and conformity aggregation according to claim 1, characterized in that, The steps for determining the lower bound of the response capacity include: for any power load resource and any set of candidate compensation control parameters, according to the adjustment direction of the current control task, performing difference processing on the corresponding baseline load forecast data and response state load forecast data to obtain multiple response capacity samples; forming a response capacity distribution based on the response capacity samples, and determining the lower bound of the response capacity corresponding to the set of candidate compensation control parameters according to the preset completion probability.
7. The intelligent control method for power compensation analysis based on demand response and conformity aggregation according to claim 1, characterized in that, The steps for determining the performance credibility of each power load resource include: extracting historical control task characteristics and historical response effect characteristics from the historical response events of each power load resource; determining the event matching weight based on the degree of matching between the historical control task characteristics and the current control task; determining the event performance evaluation value based on the historical response effect characteristics; and weighting the event performance evaluation value based on the event matching weight to determine the performance credibility of each power load resource.
8. The intelligent control method for power compensation analysis based on demand response and conformity aggregation according to claim 7, characterized in that, The steps for obtaining the schedulable response capacity include: determining a performance correction coefficient based on the performance credibility; determining a continuous correction coefficient based on the matching relationship between the response duration and the target response duration of the current control task; determining a delay correction coefficient based on the start delay and the allowed start time of the current control task; and determining a recovery correction coefficient based on the recovery risk after the response. The corresponding lower bound of the response capacity is then corrected using the performance correction coefficient, continuous correction coefficient, delay correction coefficient, and recovery correction coefficient to obtain the schedulable response capacity corresponding to each set of candidate compensation control parameters. The selection step of the target combination includes: selecting a target combination that includes the target compensation control parameters and the target schedulable response capacity from each group of candidate compensation control parameters and their corresponding schedulable response capacities, based on the degree of matching between the schedulable response capacity and the current control task, and the expected compensation cost determined by the candidate compensation control parameters.
9. The intelligent control method for power compensation analysis based on demand response and conformity aggregation according to claim 1, characterized in that, After executing the aggregated load control scheme, the method further includes: acquiring the actual response data of each power load resource; determining the response deviation based on the actual response data and the target adjustment information corresponding to the aggregated load control scheme; and classifying the response deviation into at least one of the following according to the generation stage: baseline prediction deviation, response status prediction deviation, performance execution deviation, and post-response recovery deviation; and updating at least one of the following according to the type of response deviation: the generation parameters of the baseline load prediction data, the generation parameters of the response status load prediction data, performance credibility, and post-response recovery risk.
10. A power compensation analysis and intelligent control system based on demand response and conformity aggregation, characterized in that, include: The data acquisition unit is used to acquire load control data of multiple power load resources to be aggregated, and to determine the current control task and compensation constraints. The prediction processing unit is used to set multiple sets of candidate compensation and control parameters for each power load resource, generate multiple sets of baseline load prediction data under the same operating conditions based on the load control data, and generate response state load prediction data corresponding one-to-one with the baseline load prediction data for each set of candidate compensation and control parameters. The response capacity determination unit is used to form the response capacity distribution of each power load resource under each set of candidate compensation and control parameters based on the difference between the corresponding baseline load forecast data and the response state load forecast data, and to determine the lower bound of the response capacity that satisfies the preset completion probability in each response capacity distribution. The capacity correction unit is used to determine the performance reliability of each power load resource based on historical response events that match the current control task, and, in combination with response duration, start-up delay and post-response recovery risk, correct the lower bound of the response capacity corresponding to each group of candidate compensation control parameters to obtain the dispatchable response capacity corresponding to each group of candidate compensation control parameters. The control execution unit is used to select a target combination from the candidate compensation control parameters and corresponding dispatchable response capacity of each power load resource based on the current control task and compensation constraints, and to form and execute the aggregated load control scheme with the dispatchable response capacity in the target combination as a constraint.