Virtual power plant load forecasting and demand response optimization methods and systems

Abstract

This invention provides a method and system for virtual power plant load forecasting and demand response optimization, relating to the field of power system technology. The method includes: performing multi-scale time-series decomposition of historical data to obtain a hierarchical feature set; identifying load fluctuation transmission links through cross-device correlation analysis; calibrating the load forecasting time base based on response delay time; calculating the adjustable capacity and response delay of energy equipment; dividing the equipment into multiple virtual aggregation units and establishing collaborative scheduling rules; and generating distributed demand response instructions through Nash equilibrium optimization. This invention improves the accuracy of virtual power plant load forecasting and demand response capabilities.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to a virtual power plant load prediction and demand response optimization method and system. BACKGROUND

[0002] With the rapid development of energy internet and distributed energy, as a new energy management method, virtual power plant can integrate and schedule various distributed energy resources such as dispersed renewable energy, energy storage systems and adjustable loads. Virtual power plant integrates these dispersed resources into a controllable and schedulable whole through information communication technology, participates in the electricity market and provides grid services, and plays an increasingly important role in energy transformation. Accurate load prediction and efficient demand response optimization are key links for virtual power plant to achieve economic and efficient operation, which can effectively reduce system operation cost, improve renewable energy consumption capacity, and meet the demand for stable operation of power grid.

[0003] Traditional load prediction methods usually predict each energy device as an independent individual, ignoring the complex interaction relationship and fluctuation transmission mechanism between different energy devices. In actual operation, there is significant load correlation and transmission effect between energy devices, and the load change of one device often causes the load response of other associated devices. The mutual influence between devices limits the accuracy of single device load prediction, and further affects the overall scheduling effect of virtual power plant. Existing demand response strategies mostly use centralized optimization methods, which lack fine distinction and classification management of response characteristics of different devices. Different types of energy devices have different response time delays and adjustment flexibility, and unified scheduling instructions cannot fully utilize the advantages of various devices, resulting in insufficient overall flexibility of the system, which is difficult to meet the complex and variable demand of power grid scheduling, especially in the case of high proportion of renewable energy access. The existing technology lacks an effective mechanism for collaborative optimization between devices, making it difficult to realize peak-shaving complementation and optimal resource allocation between devices. The current demand response method often cannot balance the operation cost and response benefit between devices, and there is a lack of effective negotiation and game mechanism between devices, which makes it difficult to meet the demand of power grid while ensuring the interests of each participant, resulting in low enthusiasm for demand response and limited scheduling effect. SUMMARY

[0004] The virtual power plant load prediction and demand response optimization method and system provided by the embodiments of the present application can solve the problems in the prior art.

[0005] In a first aspect, the present application provides a virtual power plant load prediction and demand response optimization method, comprising:

[0006] performing multi-scale time series decomposition on historical operation data and user-side load data of the virtual power plant to obtain a hierarchical feature set;

[0007] perform cross-device correlation analysis on each feature in the hierarchical feature set, identify inter-device load fluctuation transmission links between different energy devices in the load change time sequence, and perform device-specific calibration on the time reference of load prediction according to the response delay time of each node device in the inter-device load fluctuation transmission link, to obtain calibrated features, and generate load prediction values of each energy device based on the calibrated features;

[0008] According to the deviation between the load prediction value and the grid side scheduling demand data of the virtual power plant, combined with the current running state of the energy device, the adjustable capacity and response time delay of each energy device are calculated;

[0009] According to the response time delay and the adjustable capacity, the energy devices are divided into a plurality of virtual aggregation units, the cooperative scheduling rules between devices are established for each virtual aggregation unit, and the Nash equilibrium solution is iteratively sought between the energy devices in the virtual aggregation unit, the time sequence peak shifting and output complementation of each energy device are adjusted, and the distributed demand response scheduling instruction of the virtual power plant is generated.

[0010] The historical operation data and user side load data of the virtual power plant are subjected to multi-scale time sequence decomposition, and a hierarchical feature set is obtained, including:

[0011] According to the time dimension and the load amplitude dimension, a two-dimensional statistical space is constructed, in the two-dimensional statistical space, the time dimension is divided into a plurality of time intervals, the load amplitude dimension is divided into a plurality of load intervals, the frequency of each time interval and each load interval combination is counted, and a load-time joint probability distribution matrix is obtained;

[0012] The load-time joint probability distribution matrix is respectively edge-summed along the time dimension and the load dimension, to obtain the edge probability distribution of the time dimension and the edge probability distribution of the load amplitude, and to calculate the information entropy, to obtain the time concentration index and the load fluctuation uncertainty index; The mutual information of the load-time joint probability distribution matrix is calculated as the coupling strength index of load and time;

[0013] According to the numerical range of the load fluctuation uncertainty index, the time concentration index and the coupling strength index, the features are classified, to obtain fluctuation features, periodic features and trend features, and the hierarchical classification is performed according to the time span of the data segment, to obtain the hierarchical feature set.

[0014] Perform cross-device correlation analysis on each feature in the hierarchical feature set, identify inter-device load fluctuation transmission links between different energy devices in the load change time sequence, including:

[0015] extracting load change features of each energy equipment in the same time span level from the hierarchical feature set, taking the load change features of each energy equipment in the energy equipment pairing as driving sequence and response sequence respectively; calculating cross-correlation coefficients of the driving sequence and the response sequence at different time lag steps, identifying an optimal lag step that makes the cross-correlation coefficient reach an extreme value, and taking the optimal lag step and the extreme value of the cross-correlation coefficient as a response delay time and an association strength of the energy equipment pairing respectively;

[0016] constructing the response delay time and the association strength of all energy equipment pairings in the hierarchical feature set into an association strength matrix and a response delay matrix; retaining device pairings with association strength exceeding a preset association threshold in the association strength matrix as device nodes with coupling relationship, and determining causal directions between device nodes according to response delay times in the response delay matrix;

[0017] based on the causal directions, establishing directed edges between the device nodes with coupling relationship to obtain a directed graph structure, and extracting all directed edge paths in the directed graph structure to obtain the inter-device load fluctuation transmission link.

[0018] According to the response delay time of each node device in the inter-device load fluctuation transmission link, the time base of load prediction is calibrated device by device to obtain calibrated features, and based on the calibrated features, load prediction values of each energy equipment are generated, including:

[0019] In the inter-device load fluctuation transmission link, all response delay times in each upstream transmission path of each node device are accumulated to obtain path cumulative delay time, and the mean value of path cumulative delay times of multiple upstream transmission paths is calculated to obtain the comprehensive response delay of the node device.

[0020] A reference sequence with the comprehensive response delay as the time window length is constructed, a real-time load sequence of the energy equipment before the current prediction time is extracted, and the optimal time alignment path is determined by minimizing the dynamic time warping distance between the reference sequence and the real-time load sequence. According to the optimal time alignment path, each time point in the real-time load sequence is nonlinearly time-mapped and spliced with the load change features of all energy equipment to form the calibrated features.

[0021] The calibrated feature is constructed as a time dimension feature matrix and subjected to causal convolution processing, the convolution kernel is slid along the time dimension, and the time sequence dependent features of different time scales are extracted through layer-by-layer convolution; the residual connection is performed on each layer of time sequence dependent features, the attention scores between each time step feature vectors are calculated, the time step feature vectors are combined according to the attention scores, and the load prediction value of each energy equipment is generated through linear transformation using a fully connected layer.

[0022] According to the deviation between the load prediction value and the grid side scheduling demand data of the virtual power plant, combined with the current operating state of the energy equipment, the adjustable capacity and response time delay of each energy equipment are calculated, including:

[0023] According to the scheduling update period of the virtual power plant, the grid side scheduling demand data is divided into multiple scheduling windows, the power demand change gradient in the scheduling window is calculated and the predicted power change gradient of the corresponding load prediction value is subjected to difference operation, and the gradient deviation of each scheduling window is obtained;

[0024] The current power state and operating constraint parameters of the energy equipment are extracted from the current operating state, and a state space constraint set is constructed; in the state space constraint set, the upper boundary and the lower boundary of the state space are searched from the current power state along the power increase direction and the power decrease direction, the power distance from the current power state to the upper boundary and the lower boundary of the state space is calculated respectively and adjusted according to the gradient deviation, and the upward adjustable capacity and the downward adjustable capacity are obtained;

[0025] The response delay time of the energy equipment is extracted from the inter-device load fluctuation transmission link; according to the power ramping rate in the operating constraint parameter, the adjustment time required for the energy equipment to adjust from the current operating power to the target scheduling power is calculated, and the response delay time is superimposed on the adjustment time to obtain the response time delay of the energy equipment.

[0026] The energy equipment is divided into multiple virtual aggregation units according to the response time delay and the adjustable capacity, and the inter-device cooperative scheduling rules are established for each virtual aggregation unit, including:

[0027] A two-dimensional feature space is constructed according to the response time delay and the adjustable capacity, the two-dimensional feature space is divided into multiple grid regions by setting multiple response time delay demarcation values on the response time delay and multiple adjustable capacity demarcation values on the adjustable capacity, and all energy equipment in each grid region is determined as a virtual aggregation unit;

[0028] In the virtual aggregation unit, the energy devices are sorted in ascending order of response latency and descending order of adjustable capacity. Based on the sorting results, a device scheduling priority sequence for the virtual aggregation unit is established. The position index values ​​of the energy devices in the device scheduling priority sequence are counted, the total adjustable capacity of all energy devices in the virtual aggregation unit is calculated, and the ratio of the adjustable capacity to the total adjustable capacity is used as the capacity allocation ratio. The position index values ​​are associated with the capacity allocation ratio and stored to establish a capacity allocation mapping table for the virtual aggregation unit.

[0029] The device scheduling priority sequence and the capacity allocation mapping table are combined to form the collaborative scheduling rule.

[0030] The Nash equilibrium solution is iteratively sought among the energy devices within the virtual aggregation unit, and the timing shifting and output complementarity of each energy device are adjusted as follows:

[0031] In each iteration, the current response start time and current response output level of each energy device are obtained. The response time overlap between energy devices is calculated based on the current response start time, and the output gradient similarity is calculated based on the current response output level. The response time overlap and the output gradient similarity are combined to obtain the penalty quantization value.

[0032] The virtual aggregation unit is traversed, and the response start time and response output level that minimize the penalty quantization value are selected as the response strategy of the virtual aggregation unit in the current iteration. When the change of the response strategy is lower than the preset convergence threshold, the iteration is terminated. The optimal response start time and optimal response output level are extracted from the final response strategy as the Nash equilibrium solution.

[0033] The optimal response start-up times are sequentially arranged, the start-up time intervals between adjacent energy devices are calculated, and it is verified whether the preset minimum timing interval requirement is met. If not, the start-up times of adjacent energy devices are delayed in ascending order of response priority. The output power level of the optimal response is analyzed according to the timing sequence of the optimal response start-up times, and the output gradient difference value between adjacent energy devices is calculated. If the output gradient difference value does not meet the preset gradient complementarity requirement, the output gradient of the energy devices is redistributed in descending order of adjustable margin.

[0034] A second aspect of the present invention provides a virtual power plant load forecasting and demand response optimization system, comprising:

[0035] The first unit is used to perform multi-scale time-series decomposition on the historical operation data and user-side load data of the virtual power plant to obtain a hierarchical feature set.

[0036] The second unit is used to perform cross-device correlation analysis on each feature in the hierarchical feature set, identify the load fluctuation transmission links between different energy devices in the load change time sequence, and perform device-specific calibration on the time base of load prediction based on the response delay time of each node device in the load fluctuation transmission link to obtain calibrated features, and generate load prediction values ​​for each energy device based on the calibrated features.

[0037] The third unit is used to calculate the adjustable capacity and response delay of each energy device based on the deviation between the load forecast value and the grid-side dispatch demand data of the virtual power plant, combined with the current operating status of the energy device.

[0038] The fourth unit is used to divide the energy equipment into multiple virtual aggregation units according to the response delay and the adjustable capacity, establish collaborative scheduling rules between devices for each virtual aggregation unit, and iteratively seek Nash equilibrium solutions among the energy devices within the virtual aggregation unit to adjust the timing of peak shifting and output complementarity for each energy device, and generate distributed demand response scheduling instructions for the virtual power plant.

[0039] A third aspect of the present invention,

[0040] An electronic device is provided, comprising:

[0041] processor;

[0042] Memory used to store processor-executable instructions;

[0043] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0044] Fourth aspect of the embodiments of the present invention,

[0045] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0046] The beneficial effects of this application are as follows:

[0047] This invention improves the accuracy of virtual power plant load forecasting by performing multi-scale time-series decomposition on historical operating data and user-side load data, and by calibrating the forecast benchmark based on cross-equipment correlation analysis. It effectively avoids the forecasting deviation caused by neglecting the load transfer relationship between equipment in traditional methods.

[0048] This invention categorizes and aggregates energy equipment according to response delay and adjustable capacity, and establishes collaborative scheduling rules that adapt to the characteristics of different types of equipment. This solves the problem that traditional scheduling methods do not fully consider equipment heterogeneity, and realizes a more flexible and accurate demand response strategy.

[0049] This invention performs iterative optimization within the virtual aggregation unit based on Nash equilibrium theory, enabling the timing of energy equipment to be adjusted for peak shaving and power output complementarity. This not only improves the overall system's operating efficiency and economy but also enhances the virtual power plant's responsiveness to grid-side dispatch commands, providing strong support for the safe and stable operation of the power grid. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the virtual power plant load forecasting and demand response optimization method according to an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of the multi-scale temporal decomposition and feature extraction process. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0054] Figure 1 This is a flowchart illustrating the virtual power plant load forecasting and demand response optimization method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0055] Multi-scale time-series decomposition was performed on the historical operation data of the virtual power plant and the user-side load data to obtain a hierarchical feature set.

[0056] Cross-device correlation analysis is performed on each feature in the hierarchical feature set to identify the load fluctuation transmission links between different energy devices in the load change time sequence. The time base for load prediction is calibrated by device according to the response delay time of each node device in the load fluctuation transmission link to obtain calibrated features. Load prediction values ​​for each energy device are generated based on the calibrated features.

[0057] Based on the deviation between the load forecast value and the grid-side dispatch demand data of the virtual power plant, and combined with the current operating status of the energy equipment, the adjustable capacity and response delay of each energy equipment are calculated.

[0058] The energy equipment is divided into multiple virtual aggregation units according to the response delay and the adjustable capacity. Cooperative scheduling rules between devices are established for each virtual aggregation unit. Nash equilibrium solutions are iteratively sought among the energy equipment within each virtual aggregation unit to adjust the timing of peak shifting and output complementarity for each energy equipment, thereby generating distributed demand response scheduling instructions for the virtual power plant.

[0059] In one optional implementation, the historical operating data of the virtual power plant and the user-side load data are decomposed into a multi-scale time series to obtain a hierarchical feature set including:

[0060] A two-dimensional statistical space is constructed for the historical operation data and the user-side load data according to the time dimension and the load amplitude dimension. In the two-dimensional statistical space, the time dimension is divided into multiple time intervals and the load amplitude dimension is divided into multiple load intervals. The frequency of the combination of each time interval and each load interval is counted to obtain the load-time joint probability distribution matrix.

[0061] The load-time joint probability distribution matrix is ​​marginalized and summed along the time and load dimensions respectively to obtain the marginal probability distribution of the time dimension and the marginal probability distribution of the load amplitude. The information entropy is then calculated to obtain the time concentration index and the load fluctuation uncertainty index. The mutual information of the load-time joint probability distribution matrix is ​​calculated as an index of the coupling strength between load and time.

[0062] Based on the numerical ranges of the load fluctuation uncertainty index, the time concentration index, and the coupling strength index, feature classification is performed to obtain fluctuation features, periodic features, and trend features. These features are then hierarchically categorized according to the time span of their respective data segments to obtain the hierarchical feature set.

[0063] like Figure 2 As shown, the method includes:

[0064] Historical operating data and user-side load data of the virtual power plant are extracted from the data management system of the virtual power plant. For example, the operating data of a virtual power plant over the past 365 days can be obtained, including the load value of a sampling point every 15 minutes, totaling 35,040 data points. At the same time, user-side load data can be obtained, including the electricity consumption of various types of users and their time distribution characteristics.

[0065] A two-dimensional statistical space was constructed based on the acquired historical operational data and user-side load data, categorized by time and load amplitude. In the time dimension, a 24-hour day was divided into 48 time intervals, each lasting 30 minutes. In the load amplitude dimension, based on the maximum and minimum values ​​of historical load data, the load range was divided into 20 equally spaced load intervals; for example, if the load range is 0-100MW, each load interval would be 5MW. A 48×20 load-time joint probability distribution matrix was constructed by statistically analyzing the frequency of combinations of each time interval and each load interval. Specifically, for each element (i, j) in the matrix, the number of times the load value within time interval i falls within load interval j was counted, and then divided by the total number of samples to obtain the probability value. For example, the probability that the load value falls within the 15-20MW interval during the 7:00-7:30 AM time period is 0.023.

[0066] Summing the constructed load-time joint probability distribution matrix along the load dimension yields the marginal probability distribution along the time dimension, i.e., the probability distribution for each time interval; summing along the time dimension yields the marginal probability distribution of the load amplitude, i.e., the probability distribution for each load interval. For example, for the marginal probability distribution along the time dimension, the probability of the time period from 7:00 to 7:30 AM is calculated to be 0.021; for the marginal probability distribution along the load dimension, the probability of the load value being in the 15-20 MW range is calculated to be 0.035.

[0067] The time concentration index is calculated based on the marginal probability distribution along the time dimension. Specifically, the information entropy of the marginal probability distribution along the time dimension is calculated. The lower the information entropy, the higher the time concentration, meaning the load is concentrated in certain specific time periods. For example, a calculated information entropy of 2.73 for the marginal probability distribution along the time dimension indicates that the load of the virtual power plant has a certain degree of temporal concentration. The load fluctuation uncertainty index is calculated based on the marginal probability distribution along the load amplitude. Specifically, the information entropy of the marginal probability distribution along the load amplitude is calculated. The higher the information entropy, the greater the uncertainty of load fluctuation. For example, a calculated information entropy of 3.21 for the marginal probability distribution along the load amplitude indicates that the load amplitude changes of the virtual power plant have high uncertainty.

[0068] The mutual information of the load-time joint probability distribution matrix is ​​calculated as an indicator of the coupling strength between load and time. A higher mutual information value indicates a stronger correlation between load and time. The calculation first calculates the information entropy of the joint probability distribution matrix, then sums the information entropies of the marginal probability distributions in the time and load dimensions; the difference between the two is the mutual information. For example, a mutual information value of 1.56 for the load-time joint probability distribution matrix indicates a strong correlation between load and time.

[0069] The characteristics are classified based on the numerical ranges of the calculated load fluctuation uncertainty index, time concentration index, and coupling strength index. These characteristics are categorized into different classes according to pre-set thresholds. For example, a load fluctuation uncertainty index greater than 3.0 is considered high volatility; a time concentration index less than 2.5 is considered high periodicity; and a coupling strength index greater than 1.5 is considered high trend. In this example, the calculation results show that the virtual power plant exhibits both high volatility and high trend characteristics, but its periodicity is not significant.

[0070] Based on the time span of the data segment, the data is stratified and categorized to obtain a stratified feature set. Features can be stratified according to different time scales, such as hourly, daily, weekly, and monthly. For example, daily load data is divided into 24 hours; weekly load data into 7 days; and monthly load data into 30 days. This stratification method allows for the analysis of load characteristics at different time scales. For instance, at the hourly scale, the virtual power plant exhibits significant volatility, with a volatility uncertainty index of 3.21; at the daily scale, it exhibits obvious periodicity, with a time concentration index of 2.3; and at the monthly scale, it exhibits obvious trend characteristics, with a coupling strength index of 1.8.

[0071] Through the above steps, the multi-scale time-series decomposition of the historical operation data of the virtual power plant and the user-side load data was completed, resulting in a hierarchical feature set containing volatility, periodicity, and trend characteristics, which provides a data foundation for subsequent load forecasting and optimized scheduling.

[0072] In one optional implementation, cross-device correlation analysis is performed on each feature in the hierarchical feature set to identify the inter-device load fluctuation transmission links between different energy devices in the load change time sequence, including:

[0073] The load change characteristics of each energy device under the same time span level are extracted from the hierarchical feature set. The load change characteristics of each energy device in the energy device pairing are respectively used as the driving sequence and the response sequence. The cross-correlation coefficient of the driving sequence and the response sequence under different time lag steps is calculated. The optimal lag step that makes the cross-correlation coefficient reach the extreme value is identified. The optimal lag step and the extreme value of the cross-correlation coefficient are respectively used as the response delay time and correlation strength of the energy device pairing.

[0074] The response delay time and the association strength of all energy device pairings in the hierarchical feature set are used to construct an association strength matrix and a response delay matrix; device pairings in the association strength matrix whose association strength exceeds a preset association threshold are retained as device nodes with a coupling relationship, and the causal direction between device nodes is determined according to the response delay time in the response delay matrix.

[0075] Based on the causal direction, directed edges are established between the device nodes with coupling relationships to obtain a directed graph structure. All directed edge paths in the directed graph structure are extracted to obtain the load fluctuation transmission link between the devices.

[0076] Load variation characteristics of each energy device within the same time span are extracted from a hierarchical feature set. These characteristics include short-term (hourly), medium-term (daily), and long-term (weekly) load variation patterns. Taking an energy network in an industrial park as an example, assuming there are three energy devices: electricity load A, heat load B, and cooling load C, their load data within the same time span are extracted to form time series data. For example, in an industrial park, 72 consecutive hours of load data for these three types of devices were acquired, with sampling occurring hourly, resulting in three time series of length 72.

[0077] For each pair of energy devices, the load change characteristics of one device are considered as the driving sequence, and the load change characteristics of the other device are considered as the response sequence. Taking devices A and B as examples, the load change time series of A is used as the driving sequence, and the load change time series of B is used as the response sequence. For example, the driving sequence of A is [100, 105, 110, 108, 106...], and the response sequence of B is [50, 52, 55, 58, 57...].

[0078] The cross-correlation coefficients between the driving and response sequences are calculated at different time lag steps. These cross-correlation coefficients measure the similarity between the two time series, taking into account the time lag factor. The cross-correlation coefficients are calculated by sliding the response sequence relative to the driving sequence at different lag steps. For example, when the lag step is 0, the correlation coefficient between the driving and response sequences is calculated directly; when the lag step is 1, the response sequence is shifted forward by one time unit before calculating the correlation coefficient. Assuming that the cross-correlation coefficients obtained within a lag step range of 0 to 5 hours are [0.65, 0.78, 0.82, 0.75, 0.68, 0.60], the results are as follows.

[0079] The optimal lag step size that maximizes the cross-correlation coefficient is identified. This lag step size is used as the response delay time for device pairing, and the corresponding extreme value of the cross-correlation coefficient is used as the correlation strength. In the example above, the cross-correlation coefficient reaches its maximum value of 0.82 at a lag step size of 2 hours. Therefore, the response delay time between A and B is determined to be 2 hours, and the correlation strength is 0.82. This indicates that the load change of device B typically lags behind that of device A by approximately 2 hours, and there is a strong correlation between the two.

[0080] Repeat the above process for all device pairings to construct the association strength matrix and response delay matrix. Assuming there are n devices, this results in two n×n matrices. In this example, for three devices A, B, and C, two 3×3 matrices are constructed. The association strength matrix is ​​[[1.0, 0.82, 0.45], [0.35, 1.0, 0.75], [0.30, 0.40, 1.0]], and the response delay matrix is ​​[[0, 2, 1], [1, 0, 3], [2, 1, 0]].

[0081] Based on a preset association threshold, such as 0.6, device pairs in the association strength matrix are filtered, and pairs with association strength exceeding the threshold are retained as device nodes with a coupling relationship. In the above matrix, the association strength of pairs A→B (association strength 0.82) and B→C (association strength 0.75) is identified as exceeding the threshold of 0.6, indicating that there is a significant coupling relationship between these device pairs.

[0082] The causal direction between device nodes is determined based on the response delay time in the response delay matrix. The response delay time represents the time required from a load change in the driving device to a noticeable response from the responding device. For the A→B pairing, the response delay time is 2 hours, indicating that the load change in B lags behind that in A; for the B→C pairing, the response delay time is 3 hours, indicating that the load change in C lags behind that in B.

[0083] Based on a defined causal direction, directed edges are established between coupled device nodes to construct a directed graph structure. In this example, a directed graph containing three nodes (A, B, and C) and two directed edges (A→B and B→C) is created. This directed graph represents the load fluctuation transmission relationship between energy devices.

[0084] By extracting all directed edge paths from the constructed directed graph structure, the load fluctuation transmission links between devices are obtained. In this example, the transmission link A→B→C is identified, indicating that the load fluctuation is first transmitted from device A to device B, and then from device B to device C. This transmission link information is of great value for understanding the cascading effects in energy systems and optimizing energy management.

[0085] The above methods can automatically identify the load fluctuation transmission path in the energy equipment network, providing a key basis for the optimized control and preventive maintenance of the energy system.

[0086] In one optional implementation, the time base for load prediction is calibrated by device based on the response delay time of each node device in the inter-device load fluctuation transmission link to obtain calibrated features. The load prediction value for each energy device is then generated based on these calibrated features, including:

[0087] In the load fluctuation transmission link between the devices, the response delay time of each upstream transmission path of each node device is accumulated to obtain the path cumulative delay time. The average of the path cumulative delay time of multiple upstream transmission paths is calculated to obtain the comprehensive response delay of the node device.

[0088] A reference sequence with the comprehensive response delay as the time window length is constructed. The real-time load sequence of the energy equipment before the current prediction time is extracted. The optimal time alignment path is determined by minimizing the dynamic time warping distance between the reference sequence and the real-time load sequence. According to the optimal time alignment path, nonlinear time mapping is performed on each time point in the real-time load sequence and it is spliced ​​with the load change characteristics of all energy equipment to form the calibrated features.

[0089] The calibrated features are constructed into a time-dimensional feature matrix and subjected to causal convolution. The convolution kernel is slid along the time dimension, and time-dependent features at different time scales are extracted through layer-by-layer convolution. Residual connections are made between the time-dependent features of each layer, and attention scores are calculated between the feature vectors of each time step. The time-step feature vectors are combined according to the attention scores and linearly transformed using a fully connected layer to generate the load prediction values ​​of each energy device.

[0090] Analyze the response delay time of each node device in the load fluctuation transmission link between devices. For example, in the above-mentioned combined cooling, heating and power system, when the user's electricity load increases, the output of the gas turbine will increase first, followed by the steam output of the waste heat boiler, which in turn affects the cooling capacity of the absorption chiller and the power generation of the steam turbine.

[0091] For each node device in the transmission link, identify all its upstream transmission paths. For example, for an absorption chiller, its upstream transmission paths include two paths: "gas turbine → waste heat boiler → absorption chiller" and "gas turbine → absorption chiller". For each path, sum the response delay times of all devices in the path to obtain the path cumulative delay time. For example, if the gas turbine response delay is 5 minutes and the waste heat boiler response delay is 8 minutes, then the cumulative delay of the first path is 13 minutes. Calculate the average of the cumulative delay times of all upstream transmission paths as the overall response delay of the node device. In this example, assuming the cumulative delays of the two paths are 13 minutes and 7 minutes respectively, the overall response delay of the absorption chiller is 10 minutes.

[0092] Based on the calculated overall response delay, a reference sequence is constructed, with the time window length of the reference sequence equal to the overall response delay. For example, for an absorption chiller with an overall response delay of 10 minutes, a reference sequence of 10 time points is constructed, which can be an ideal load response curve or a historical typical load change pattern.

[0093] Extract the real-time load sequence of energy devices up to the current forecast time. For example, extract the real-time load data of the absorption chiller for the most recent 30 time points to form a real-time load sequence. Calculate the distance matrix between the reference sequence and the real-time load sequence using a dynamic time warping algorithm. In this matrix, each element represents the distance between the i-th point in the reference sequence and the j-th point in the real-time load sequence. Find the path with the minimum cumulative distance value, i.e., the optimal time alignment path, using a dynamic programming algorithm. For example, if time points 5, 8, 12, 15, 19, 22, 25, 27, 28, and 30 in the real-time load sequence correspond to 10 points in the reference sequence, then these 10 time points constitute the optimal time alignment path.

[0094] Based on the optimal time alignment path, nonlinear time mapping is performed on each time point in the real-time load sequence. The time points on the optimal path in the real-time load sequence are extracted and concatenated with the load change characteristics of all energy devices to form calibrated features. These features can include multi-dimensional information such as load values, load change rates, and equipment operating status.

[0095] To generate load forecasts based on calibrated features, the calibrated features are constructed into a time-dimensional feature matrix. Assuming there are n energy devices, and each device extracts features at m time points, with k feature dimensions at each time point, the constructed feature matrix has dimensions n×m×k.

[0096] Causal convolution is performed on the feature matrix, using multiple convolutional kernels of different sizes, such as kernels with widths of 3, 5, and 7. These kernels slide along the time dimension to perform the convolution operation, capturing load variation patterns at different time scales. For example, a kernel with a width of 3 focuses on short-term load fluctuations, while a kernel with a width of 7 captures longer-term load trends. Through layer-by-layer convolution, temporal dependency features at different time scales are extracted from the calibrated features.

[0097] For the extracted temporal dependency features from each layer, residual connections are used for feature fusion. Residual connections can effectively alleviate the gradient vanishing problem in deep networks and ensure effective information transfer. Attention scores are calculated between feature vectors at each time step to determine the importance of different time points to the prediction results. For example, if recent load changes have a greater impact on future loads, the attention scores at recent time steps will be higher.

[0098] Based on the calculated attention scores, the feature vectors at each time step are weighted and combined. This weighted combination highlights the influence of important time points and weakens the interference of irrelevant time points. Finally, a fully connected layer is used to perform a linear transformation on the combined features to generate the load forecast values ​​for each energy device.

[0099] By using the above methods, the time delay characteristics of load fluctuation transmission between devices can be taken into account, and the load forecast time base of different devices can be accurately calibrated, thereby improving the accuracy and reliability of overall load forecasting and providing strong support for the optimized scheduling of energy systems.

[0100] In one optional implementation, the adjustable capacity and response delay of each energy device are calculated based on the deviation between the load forecast value and the grid-side dispatch demand data of the virtual power plant, combined with the current operating status of the energy devices, including:

[0101] According to the scheduling update cycle of the virtual power plant, the grid-side scheduling demand data is divided into multiple scheduling windows. The power demand change gradient within the scheduling window is calculated and differentially calculated with the predicted power change gradient of the corresponding load forecast value to obtain the gradient deviation of each scheduling window.

[0102] Extract the current power state and operating constraint parameters of the energy equipment from the current operating state to construct a state space constraint set; in the state space constraint set, starting from the current power state, search the upper and lower boundaries of the state space along the power increase and power decrease directions, calculate the power distance from the current power state to the upper and lower boundaries of the state space respectively, and make adjustments according to the gradient deviation to obtain the upward adjustable capacity and the downward adjustable capacity;

[0103] The response delay time of the energy device is extracted from the load fluctuation transmission link between the devices; the adjustment time required for the energy device to adjust from the current operating power to the target scheduling power is calculated according to the power ramp rate in the operating constraint parameters; the response delay time is superimposed with the adjustment time to obtain the response delay of the energy device.

[0104] The grid-side dispatch demand data is divided into multiple dispatch windows according to the dispatch update cycle of the virtual power plant. For example, if the dispatch update cycle is 15 minutes, the dispatch demand data for the next 4 hours is divided into 16 dispatch windows. For each dispatch window, its power demand change gradient is calculated, that is, the change in power per unit time. Assuming that in the 5th dispatch window, the grid dispatch demand increases from 25MW to 27MW, the power demand change gradient is (27-25) / 15=0.133MW / minute. Similarly, the predicted power change gradient of the load forecast value in the corresponding time period is calculated. For example, if the load forecast increases from 23MW to 24MW in this time period, the predicted power change gradient is (24-23) / 15=0.067MW / minute. By performing a difference operation on the two, the gradient deviation of the dispatch window is obtained as 0.133-0.067=0.066MW / minute, indicating that the actual demand is increasing at a faster rate than the forecast.

[0105] The current power state and operating constraint parameters are extracted from the current operating status of energy equipment. The current power state refers to the current output power level of the equipment, such as the current output power of a photovoltaic power station being 8MW. Operating constraint parameters include the maximum / minimum power limits and ramp rate limits of the equipment. For example, the power range of an energy storage power station is 0-5MW, and the charging / discharging ramp rate limit is 0.5MW / minute; the power range of a gas turbine is 2-10MW, and the ramp rate limit is 0.2MW / minute.

[0106] Based on these parameters, a set of state-space constraints is constructed, which represents the range of power states within which the equipment can operate. Within this set of constraints, starting from the current power state, the upper and lower boundaries of the state space are searched along the power increase and decrease directions. For example, a photovoltaic power station currently outputs 8MW, with a maximum output capacity of 12MW. Considering weather factors, it can actually be increased to 10MW, so the upper boundary is 10MW. The minimum output is 0MW, but considering curtailment restrictions, it can actually be reduced to a minimum of 2MW, so the lower boundary is 2MW.

[0107] Calculate the power distance from the current power state to the upper and lower boundaries of the state space, i.e., the upward adjustable capacity is 10-8=2MW, and the downward adjustable capacity is 8-2=6MW. Adjustments are made based on the previously calculated gradient deviation. If the gradient deviation is positive, it indicates a faster power increase is needed, requiring an increase in the upward adjustable capacity; if the gradient deviation is negative, the downward adjustable capacity needs to be increased. Assuming a gradient deviation of 0.066MW / minute and considering a 15-minute scheduling cycle, the upward adjustable capacity can be adjusted to 2+0.066×15≈3MW, while the downward adjustable capacity remains unchanged at 6MW.

[0108] The response delay time of energy equipment is extracted from the load fluctuation transmission link between devices. The response delay time refers to the time delay from receiving the dispatch command to starting the dispatch. For example, the response delay time of an energy storage power station is 30 seconds, and the response delay time of a gas turbine is 2 minutes. Based on the power ramp rate in the operating constraint parameters, the adjustment time required for the energy equipment to adjust from the current operating power to the target dispatch power is calculated. For example, if an energy storage power station is currently outputting 1MW and needs to adjust to 4MW, with a ramp rate of 0.5MW / minute, then the adjustment time is (4-1) / 0.5=6 minutes.

[0109] The response delay time is obtained by adding the response delay time to the adjustment time. For the energy storage power station mentioned above, the response delay is 30 seconds + 6 minutes = 6.5 minutes, indicating that it takes 6.5 minutes from issuing the dispatch command to reaching the target dispatch power. This response delay data is crucial for the dispatch optimization of virtual power plants and can be used to evaluate the equipment's ability to participate in ancillary services such as grid frequency regulation and peak shaving.

[0110] By calculating the adjustable capacity and response delay data, the virtual power plant can more accurately assess the adjustment capabilities of each energy device, thereby achieving optimized scheduling of energy devices while meeting the grid-side dispatching requirements, and improving the system's economy and stability.

[0111] In one optional implementation, the energy equipment is divided into multiple virtual aggregation units according to the response delay and the adjustable capacity, and collaborative scheduling rules between devices are established for each virtual aggregation unit, including:

[0112] A two-dimensional feature space is constructed using the response delay and the adjustable capacity. By setting multiple response delay boundary values ​​and multiple adjustable capacity boundary values ​​on the adjustable capacity, the two-dimensional feature space is divided into multiple grid regions. All energy devices in each grid region are defined as a virtual aggregation unit.

[0113] In the virtual aggregation unit, the energy devices are sorted in ascending order of response latency and descending order of adjustable capacity. Based on the sorting results, a device scheduling priority sequence for the virtual aggregation unit is established. The position index values ​​of the energy devices in the device scheduling priority sequence are counted, the total adjustable capacity of all energy devices in the virtual aggregation unit is calculated, and the ratio of the adjustable capacity to the total adjustable capacity is used as the capacity allocation ratio. The position index values ​​are associated with the capacity allocation ratio and stored to establish a capacity allocation mapping table for the virtual aggregation unit.

[0114] The device scheduling priority sequence and the capacity allocation mapping table are combined to form the collaborative scheduling rule.

[0115] A two-dimensional feature space is constructed using response latency and adjustable capacity. Specifically, the horizontal axis represents response latency, which can be in seconds or minutes; the vertical axis represents adjustable capacity, which can be in kilowatts or megawatts. For example, for a specific application scenario, the response latency can be set to a range of 0 to 300 seconds, and the adjustable capacity can be set to a range of 0 to 1000 kilowatts.

[0116] After constructing the two-dimensional feature space, multiple response delay thresholds are set on the response delay axis and multiple adjustable capacity thresholds are set on the adjustable capacity axis, thereby dividing the two-dimensional feature space into multiple grid regions. For example, thresholds of 30 seconds, 60 seconds, 120 seconds, and 180 seconds can be set on the response delay axis, and thresholds of 100 kW, 200 kW, 500 kW, and 800 kW can be set on the adjustable capacity axis. Through these thresholds, the two-dimensional feature space is divided into 5×5=25 grid regions.

[0117] Each grid region contains energy devices with similar response latency and adjustable capacity characteristics. For example, all energy devices within a grid region (0-30 seconds, 500-800 kW) have response latency between 0 and 30 seconds and adjustable capacity between 500 and 800 kW. All energy devices within each grid region are defined as a virtual aggregation unit. Assuming there are 100 energy devices distributed across these 25 grid regions, with each grid region containing a different number of devices, thus forming 25 virtual aggregation units.

[0118] For each virtual aggregation unit, a scheduling priority sequence for its internal devices needs to be established. Specifically, the energy devices within the virtual aggregation unit are sorted in ascending order of response latency and descending order of adjustable capacity. For example, for energy devices in a grid area (0-30 seconds, 500-800 kW), assume there are devices A (10 seconds, 700 kW), B (15 seconds, 650 kW), C (5 seconds, 600 kW), D (5 seconds, 750 kW), and E (20 seconds, 780 kW). According to the sorting rules, they are first sorted in ascending order of response latency as follows: Device C (5 seconds, 600 kW), Device D (5 seconds, 750 kW), Device A (10 seconds, 700 kW), Device B (15 seconds, 650 kW), and Device E (20 seconds, 780 kW). For devices C and D with the same response latency, they are then sorted in descending order of adjustable capacity. Therefore, the final sorting result is: Device D (5 seconds, 750 kW), Device C (5 seconds, 600 kW), Device A (10 seconds, 700 kW), Device B (15 seconds, 650 kW), Device E (20 seconds, 780 kW). This sorting result is the device scheduling priority sequence for this virtual aggregation unit.

[0119] Based on the established equipment scheduling priority sequence, the position index value of the energy equipment in the equipment scheduling priority sequence is counted. Taking the above example, the position index value of equipment D is 1, the position index value of equipment C is 2, the position index value of equipment A is 3, the position index value of equipment B is 4, and the position index value of equipment E is 5.

[0120] Then, the total adjustable capacity of all energy devices within the virtual aggregation unit is calculated. Continuing with the example above, the total adjustable capacity is 750+600+700+650+780=3480 kilowatts.

[0121] For each energy device, the ratio of its adjustable capacity to the total adjustable capacity is used as the capacity allocation ratio. For example, the capacity allocation ratio for device D is 750÷3480≈0.216, the capacity allocation ratio for device C is 600÷3480≈0.172, the capacity allocation ratio for device A is 700÷3480≈0.201, the capacity allocation ratio for device B is 650÷3480≈0.187, and the capacity allocation ratio for device E is 780÷3480≈0.224.

[0122] The location index values ​​are associated with the capacity allocation ratios to establish a capacity allocation mapping table for the virtual aggregation unit. For example, for the virtual aggregation unit mentioned above, its capacity allocation mapping table is: {1: 0.216, 2: 0.172, 3: 0.201, 4: 0.187, 5: 0.224}. This mapping table indicates that when the virtual aggregation unit needs to be called for a response, the device with a location index value of 1 (i.e., device D) should be allocated 21.6% of the total scheduling capacity, the device with a location index value of 2 (i.e., device C) should be allocated 17.2% of the total scheduling capacity, and so on.

[0123] The device scheduling priority sequence and capacity allocation mapping table are combined to form a collaborative scheduling rule. For the above virtual aggregation unit, its collaborative scheduling rule can be expressed as: {priority sequence: [device D, device C, device A, device B, device E], capacity allocation mapping: {1: 0.216, 2: 0.172, 3: 0.201, 4: 0.187, 5: 0.224}}.

[0124] When energy equipment needs to be dispatched in response, a suitable virtual aggregation unit is first determined based on the response requirements (such as the required response time and adjustment capacity). For example, if a response is required within 20 seconds and 600 kW of capacity needs to be adjusted, a virtual aggregation unit with a response delay of no more than 20 seconds and a total adjustable capacity of no less than 600 kW is selected. Then, according to the cooperative scheduling rules of the selected virtual aggregation unit, the equipment is called in order of equipment scheduling priority, and an appropriate adjustment capacity is allocated to each equipment according to the capacity allocation mapping table. For example, if the total capacity to be adjusted is 600 kW, according to the above capacity allocation mapping table, equipment D will be assigned an adjustment task of 600 × 0.216 ≈ 129.6 kW, equipment C will be assigned an adjustment task of 600 × 0.172 ≈ 103.2 kW, and so on.

[0125] In this way, energy equipment is effectively organized and managed, enabling rapid and efficient scheduling and response based on actual needs, thus improving the flexibility and reliability of the energy system.

[0126] In one optional implementation, the Nash equilibrium solution is iteratively sought among the energy devices within the virtual aggregation unit to adjust the timing shifting and output complementarity for each energy device, including:

[0127] In each iteration, the current response start time and current response output level of each energy device are obtained. The response time overlap between energy devices is calculated based on the current response start time, and the output gradient similarity is calculated based on the current response output level. The response time overlap and the output gradient similarity are combined to obtain the penalty quantization value.

[0128] The virtual aggregation unit is traversed, and the response start time and response output level that minimize the penalty quantization value are selected as the response strategy of the virtual aggregation unit in the current iteration. When the change of the response strategy is lower than the preset convergence threshold, the iteration is terminated. The optimal response start time and optimal response output level are extracted from the final response strategy as the Nash equilibrium solution.

[0129] The optimal response start-up times are sequentially arranged, the start-up time intervals between adjacent energy devices are calculated, and it is verified whether the preset minimum timing interval requirement is met. If not, the start-up times of adjacent energy devices are delayed in ascending order of response priority. The output power level of the optimal response is analyzed according to the timing sequence of the optimal response start-up times, and the output gradient difference value between adjacent energy devices is calculated. If the output gradient difference value does not meet the preset gradient complementarity requirement, the output gradient of the energy devices is redistributed in descending order of adjustable margin.

[0130] This embodiment provides a method for timing staggering and output complementarity of energy devices. By iteratively seeking a Nash equilibrium solution among the energy devices within a virtual aggregation unit, timing staggering and output complementarity are adjusted for each energy device.

[0131] In practical implementation, the virtual aggregation unit comprises multiple energy devices, such as photovoltaic power generation equipment, wind power generation equipment, and energy storage batteries, each with unique response and output characteristics. To achieve peak shaving and output complementarity, it is necessary to coordinate the response start-up time and output level of these devices to optimize overall energy utilization.

[0132] In the process of iteratively seeking the Nash equilibrium solution, in the first iteration, based on the initial state of each energy device, the initial response start time and the initial response output level are set. For example, for 5 energy devices contained in a virtual aggregation unit, the initial response start time can be set to [0 minutes, 15 minutes, 30 minutes, 45 minutes, 60 minutes], and the initial response output level can be set to 50% of the rated power of each device.

[0133] In each iteration, the current response start-up time and current response output level of each energy device are obtained. Assuming that after the first iteration, the response start-up times of the five devices become [5 minutes, 18 minutes, 32 minutes, 48 ​​minutes, 65 minutes], and the response output levels become [55%, 48%, 52%, 58%, 45% of rated power]. Based on this data, the response time overlap between energy devices is calculated. Response time overlap represents the degree of overlap in the response times of different devices, which can be quantified by calculating the ratio between the time difference between the start-up times of any two devices and the preset minimum timing interval requirement. For example, if the start-up time difference between device 1 and device 2 is 13 minutes (18-5), and the preset minimum timing interval is 15 minutes, then their overlap is (15-13) / 15 = 0.133, indicating a 13.3% time overlap.

[0134] The output gradient similarity is calculated based on the current response output level. The output gradient represents the rate of change of equipment output over time, and the output gradient similarity measures the degree of similarity in the output change trends of different devices. For example, if the output gradient of device 1 is +2.5% / min and the output gradient of device 2 is +2.1% / min, the gradient similarity between the two can be calculated as the ratio of the dot product of their gradient directions to the product of their respective absolute gradient values. The result is 0.84, indicating that the output change trends of the two devices are 84% similar.

[0135] The penalty quantization value is obtained by combining the response time overlap and the output gradient similarity. The combination method can be a weighted sum, i.e., penalty quantization value = 0.6 × response time overlap + 0.4 × output gradient similarity. The weights can be adjusted according to the actual application requirements. In this example, more emphasis is placed on timing staggering, so the time overlap has a higher weight. For the above example, the penalty quantization value is 0.6 × 0.133 + 0.4 × 0.84 = 0.4159.

[0136] The algorithm iterates through all response strategy combinations within the virtual aggregation unit. Each combination contains specific response start times and response output levels. For example, it might try adjusting the start time of device 1 to 8 minutes and the output level to 53%, calculating a new penalty quantization value of 0.3985, which is lower than the current value. Then, it might try adjusting the start time of device 2 to 22 minutes and the output level to 45%, calculating a new penalty quantization value of 0.3782, further reducing the penalty. Through this comprehensive iteration, the response start time and response output level that minimizes the penalty quantization value are selected as the response strategy for the virtual aggregation unit in the current iteration round.

[0137] The iteration terminates when the change in the response strategy falls below a preset convergence threshold. This change can be defined as the average percentage change in the response initiation time and response output level over two consecutive iterations. For example, if the preset convergence threshold is 2%, and the change in the response strategy is 1.8% after the fifth iteration, the iteration terminates. The optimal response initiation time and optimal response output level are extracted from the final response strategy as the Nash equilibrium solution.

[0138] After obtaining the Nash equilibrium solution, the optimal response start times are sequentially arranged, i.e., the start times of each device are arranged in chronological order. Assuming the arranged start times are [5 minutes, 18 minutes, 32 minutes, 48 ​​minutes, 65 minutes], the start time intervals between adjacent energy devices are calculated to be 13 minutes, 14 minutes, 16 minutes, and 17 minutes, respectively. These time intervals are verified to meet the preset minimum time interval requirement, such as 15 minutes. In this example, the time intervals of the first two groups of devices do not meet the requirement. Therefore, the start times of adjacent energy devices are delayed in ascending order of response priority. Assuming the response priorities of devices 1, 2, 3, 4, and 5 are 3, 5, 1, 4, and 2, respectively, the priority ascending order is 3, 5, 1, 4, and 2. The start time of the lower-priority device 2 is delayed from 18 minutes to 20 minutes, making the interval 15 minutes. Similarly, the start time of device 3 is delayed from 32 minutes to 35 minutes.

[0139] For the optimal response output level, output gradient analysis is performed according to the time sequence of the adjusted optimal response start-up time. The output gradient represents the change in output per unit time, which can be calculated by dividing the output difference between adjacent time points by the time difference. The output gradient difference between adjacent energy devices is calculated. For example, the output gradient difference between device 1 and device 2 is |2.5%-2.1%|=0.4% / minute. If the preset gradient complementarity requirement is that the output gradient difference is not less than 1% / minute, then device 1 and device 2 do not meet the requirement. The output gradient is redistributed to the energy devices in descending order of adjustability margin. Assuming that the adjustability margins of devices 1, 2, 3, 4, and 5 are 20%, 15%, 25%, 18%, and 22% respectively, the descending order is 3, 5, 1, 4, 2. The output gradient of device 3, which has the largest adjustability margin, is adjusted first to ensure that its output gradient difference with adjacent devices meets the preset requirement.

[0140] Through the above steps, the timing of peak shaving and output complementarity of each energy device within the virtual aggregation unit are achieved, ultimately forming a Nash equilibrium solution that satisfies the individual interests of each device while meeting the overall optimization goal.

[0141] This invention relates to a virtual power plant load forecasting and demand response optimization system, the system comprising:

[0142] The first unit is used to perform multi-scale time-series decomposition on the historical operation data and user-side load data of the virtual power plant to obtain a hierarchical feature set.

[0143] The second unit is used to perform cross-device correlation analysis on each feature in the hierarchical feature set, identify the load fluctuation transmission links between different energy devices in the load change time sequence, and perform device-specific calibration on the time base of load prediction based on the response delay time of each node device in the load fluctuation transmission link to obtain calibrated features, and generate load prediction values ​​for each energy device based on the calibrated features.

[0144] The third unit is used to calculate the adjustable capacity and response delay of each energy device based on the deviation between the load forecast value and the grid-side dispatch demand data of the virtual power plant, combined with the current operating status of the energy device.

[0145] The fourth unit is used to divide the energy equipment into multiple virtual aggregation units according to the response delay and the adjustable capacity, establish collaborative scheduling rules between devices for each virtual aggregation unit, and iteratively seek Nash equilibrium solutions among the energy devices within the virtual aggregation unit to adjust the timing of peak shifting and output complementarity for each energy device, and generate distributed demand response scheduling instructions for the virtual power plant.

[0146] A third aspect of the present invention provides an electronic device, comprising:

[0147] processor;

[0148] Memory used to store processor-executable instructions;

[0149] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0150] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0151] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for virtual power plant load forecasting and demand response optimization, characterized in that, The method comprises the following steps: performing multi-scale time series decomposition on historical operation data and user-side load data of a virtual power plant to obtain a hierarchical feature set; performing cross-device correlation analysis on each feature in the hierarchical feature set to identify inter-device load fluctuation transmission links between different energy devices in the time series of load changes, and performing device-specific calibration on the time reference of load prediction according to the response delay time of each node device in the inter-device load fluctuation transmission link to obtain calibrated features, and generating load prediction values of each energy device based on the calibrated features; calculating the adjustable capacity and response time delay of each energy device according to the deviation between the load prediction values and grid-side scheduling demand data of the virtual power plant, combined with the current operation state of the energy device; dividing the energy devices into multiple virtual aggregation units according to the response time delay and the adjustable capacity, establishing a cooperative scheduling rule between devices for each virtual aggregation unit, and iteratively seeking a Nash equilibrium solution between the energy devices within the virtual aggregation unit to adjust the timing peak shifting and output complement of each energy device, and generating a distributed demand response scheduling instruction for the virtual power plant.

2. The method of claim 1, wherein, The method comprises the following steps: performing multi-scale time series decomposition on historical operation data and user-side load data of a virtual power plant to obtain a hierarchical feature set includes: constructing a two-dimensional statistical space according to the time dimension and the load amplitude dimension, dividing the time dimension into multiple time intervals and the load amplitude dimension into multiple load intervals in the two-dimensional statistical space, and calculating the frequency of each time interval and load interval combination to obtain a load-time joint probability distribution matrix; performing edge summation along the time dimension and the load dimension on the load-time joint probability distribution matrix to obtain the edge probability distribution of the time dimension and the edge probability distribution of the load amplitude, and calculating the information entropy to obtain the time concentration index and the load fluctuation uncertainty index; and calculating the mutual information of the load-time joint probability distribution matrix as the coupling strength index of load and time; 3. The method of claim 1, wherein, performing feature classification according to the numerical range of the load fluctuation uncertainty index, the time concentration index and the coupling strength index to obtain fluctuation features, periodic features and trend features, and hierarchically classifying the features according to the time span of the data segment to obtain the hierarchical feature set. The method comprises the following steps: extracting the load change features of each energy device under the same time span level from the hierarchical feature set, taking the load change features of each energy device in the energy device pairing as the driving sequence and the response sequence respectively; calculating the cross-correlation coefficient of the driving sequence and the response sequence at different time lag steps to identify the optimal lag step that makes the cross-correlation coefficient reach an extreme value, and taking the optimal lag step and the extreme value of the cross-correlation coefficient as the response delay time and the correlation strength of the energy device pairing respectively; The response delay time and the association strength of all energy equipment pairs in the layered feature set are constructed into an association strength matrix and a response delay matrix; the equipment pairs with an association strength exceeding a preset association threshold in the association strength matrix are retained as device nodes with a coupling relationship, and the causal direction between the device nodes is determined according to the response delay time in the response delay matrix; Based on the causal direction, a directed edge is established between the device nodes with the coupling relationship to obtain a directed graph structure, and all directed edge paths in the directed graph structure are extracted to obtain the inter-device load fluctuation transmission link.

4. The method of claim 1, wherein, According to the response delay time of each node device in the inter-device load fluctuation transmission link, the time reference of load prediction is calibrated for each energy equipment to obtain a calibrated feature, and based on the calibrated feature, a load prediction value of each energy equipment is generated, including: In the inter-device load fluctuation transmission link, all response delay times in each upstream transmission path of each node device are accumulated to obtain a path cumulative delay time, the mean value of the path cumulative delay times of multiple upstream transmission paths is calculated to obtain a comprehensive response delay of the node device; A reference sequence with the comprehensive response delay as the time window length is constructed, a real-time load sequence of the energy equipment before the current prediction time is extracted, and an optimal time alignment path is determined by minimizing the dynamic time warping distance between the reference sequence and the real-time load sequence; according to the optimal time alignment path, each time point in the real-time load sequence is nonlinearly time-mapped and spliced with the load change feature of all energy equipment to form the calibrated feature; The calibrated feature is constructed into a time-dimension feature matrix and subjected to causal convolution processing, the convolution kernel is slid along the time dimension, time sequence dependent features of different time scales are extracted through layer-by-layer convolution; the residual connection is performed on each layer of time sequence dependent features, the attention score between each time step feature vector is calculated, the time step feature vectors are combined according to the attention score, and a linear transformation is performed on the combined time step feature vectors using a fully connected layer to generate the load prediction value of each energy equipment.

5. The method of claim 1, wherein, According to the deviation between the load prediction value and the grid-side scheduling demand data of the virtual power plant, and in combination with the current operating state of the energy equipment, the adjustable capacity and response time delay of each energy equipment are calculated, including: The grid-side scheduling demand data is divided into multiple scheduling windows according to the scheduling update period of the virtual power plant, the power demand change gradient in the scheduling window is calculated, and the predicted power change gradient of the load prediction value is differentially operated to obtain the gradient deviation of each scheduling window; The current power state and operating constraint parameters of the energy equipment are extracted from the current operating state to construct a state space constraint set; in the state space constraint set, the upper boundary and the lower boundary of the state space are searched from the current power state along the power increase direction and the power decrease direction, the power distance from the current power state to the upper boundary and the lower boundary of the state space is calculated respectively, and the adjustable capacity upward and the adjustable capacity downward are obtained according to the gradient deviation. extracting a response delay time of the energy equipment from the inter-device load fluctuation transmission link; calculating an adjustment time required for the energy equipment to adjust from a current operation power to a target scheduling power according to a power ramping rate in the operation constraint parameter, superimposing the response delay time and the adjustment time to obtain the response time delay of the energy equipment.

6. The method of claim 1, wherein, dividing the energy equipment into a plurality of virtual aggregation units according to the response time delay and the adjustable capacity, and establishing a cooperative scheduling rule between devices for each virtual aggregation unit, including: constructing a two-dimensional feature space with the response time delay and the adjustable capacity, dividing the two-dimensional feature space into a plurality of grid regions by setting a plurality of response time delay boundary values on the response time delay and a plurality of adjustable capacity boundary values on the adjustable capacity, and determining all energy equipment in each grid region as a virtual aggregation unit; sorting the energy equipment in the virtual aggregation unit in ascending order of the response time delay and descending order of the adjustable capacity, establishing a device scheduling priority sequence of the virtual aggregation unit based on the sorting result, counting a position index value of the energy equipment in the device scheduling priority sequence, calculating a total adjustable capacity of all energy equipment in the virtual aggregation unit, taking a ratio of the adjustable capacity to the total adjustable capacity as a capacity allocation proportion, and associating and storing the position index value and the capacity allocation proportion to establish a capacity allocation mapping table of the virtual aggregation unit; combining the device scheduling priority sequence and the capacity allocation mapping table to form the cooperative scheduling rule.

7. The method of claim 1, wherein, iteratively seeking a Nash equilibrium solution among the energy equipment in the virtual aggregation unit, and adjusting the timing peak shaving and the output complement of each energy equipment, including: in each iteration, obtaining a current response start time and a current response output level of each energy equipment, calculating a response time overlap degree between energy equipment based on the current response start time, calculating an output gradient similarity based on the current response output level, and combining the response time overlap degree and the output gradient similarity to obtain a penalty quantization value; traversing the virtual aggregation unit, selecting a response start time and a response output level that minimize the penalty quantization value as a response strategy of the virtual aggregation unit in the current iteration, and terminating the iteration when a change amount of the response strategy is less than a preset convergence threshold, extracting an optimal response start time and an optimal response output level from the final response strategy as the Nash equilibrium solution; sequentially arranging the optimal response start time, calculating a start time interval between adjacent energy equipment and verifying whether it meets a preset minimum time interval requirement, delaying the time of adjacent energy equipment in ascending order of response priority if it does not meet the requirement, and performing output gradient analysis on the optimal response output level according to the time sequence order of the optimal response start time, calculating an output gradient difference value between adjacent energy equipment, and redistributing the output gradient of the energy equipment in descending order of adjustable margin if the output gradient difference value does not meet the preset gradient complement requirement.

8. A virtual power plant load forecasting and demand response optimization system for implementing the method of any one of claims 1-7, characterized in that, The method comprises the following steps: A first unit is configured to perform multi-scale time series decomposition on historical operation data of a virtual power plant and user-side load data to obtain a set of hierarchical features; A second unit is configured to perform cross-device correlation analysis on each feature in the set of hierarchical features, identify inter-device load fluctuation transmission links between different energy devices in terms of load change time series, and perform device-specific calibration on a time reference for load prediction according to response delay times of each node device in the inter-device load fluctuation transmission links to obtain calibrated features, and generate load prediction values of each energy device based on the calibrated features; A third unit is configured to calculate adjustable capacity and response time delay of each energy device according to deviations between the load prediction values and grid-side scheduling demand data of the virtual power plant, and in combination with current operation states of the energy devices; A fourth unit is configured to divide the energy devices into a plurality of virtual aggregation units according to the response time delay and the adjustable capacity, establish cooperative scheduling rules between devices for each virtual aggregation unit, and iteratively seek Nash equilibrium solutions among the energy devices within the virtual aggregation units to adjust time series peak shifting and output complementarity for each energy device, and generate distributed demand response scheduling instructions for the virtual power plant.

9. An electronic device, comprising: The method comprises the following steps: A processor; A memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 7.

Images