Virtual Power Plant Load Forecasting Methods and Systems
By constructing a ramp-up rate sequence and identifying concurrent peak voltage activation indicators, splitting load features and implementing weight amplification, and generating a multi-dimensional feature vector, the problem of peak value underestimation during the concurrent positive ramp-up period of multiple heterogeneous sub-loads in a virtual power plant is solved, thereby improving prediction accuracy and stability.
Patent Information
- Application Number
- CN202611299708.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-08-26
AI Technical Summary
In existing technologies, the concurrent positive ramp-up period of multiple heterogeneous sub-loads is averaged due to the ramp-up phase alignment effect, resulting in a structural underestimation bias in the prediction of the composite net load peak.
By constructing a ramp rate sequence, identifying concurrent peak pressure activation indicators, splitting load features into a baseline component and a ramp excess component, and combining concurrent peak pressure scores to amplify the ramp excess component with weights, a multi-dimensional feature vector is generated, which is then input into the time series prediction model and accompanied by concurrent peak pressure markers.
It improves the structural underestimation problem of the peak net load of virtual power plants, enhances the peak tracking ability of the prediction results during the concurrent positive ramp-up period of multiple heterogeneous subloads, and reduces the risk of dispatch deviation assessment.
Smart Images

Figure CN122823406B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data processing technology, specifically relating to a method and system for virtual power plant load forecasting. Background Technology
[0002] Virtual power plants aggregate heterogeneous distributed resources within their jurisdiction, such as distributed photovoltaics, wind turbines, electrochemical energy storage, electric vehicle charging stations, industrial flexible loads, and building HVAC systems, to participate in the power grid dispatch market as a unified entity. Load forecasting is the core data foundation for virtual power plants to formulate day-ahead reporting plans and participate in ancillary service bidding; its accuracy directly determines the deviation of virtual power plants' market bids and the risk of performance penalties.
[0003] Patent application CN121958817A discloses a multi-source fusion multi-energy load forecasting method for virtual power plants. The method includes constructing aggregated multi-source heterogeneous data for the virtual power plant, including historical multi-energy load data within the virtual power plant's jurisdiction, calendar and market information, and external environmental auxiliary data. The method preprocesses the multi-source heterogeneous data, performs feature filtering on the time-series feature matrix based on correlation analysis, constructs a model input set, and builds a load forecasting model. The model includes a sequentially connected input layer, several stacked spatiotemporal fusion layers, and an output layer. The spatiotemporal fusion layer includes parallel multi-scale temporal convolutional modules and graph convolutional modules. The method also includes model training and parameter optimization, load forecasting, and result output. The method achieves joint forecasting of electrical load, thermal load, and cooling load within a unified multi-scale spatiotemporal graph neural network framework.
[0004] Current mainstream virtual power plant load forecasting schemes typically employ a method of independent forecasting by source followed by weighted aggregation. Separate forecasting models are established for each type of sub-resource; for example, irradiance regression is used for photovoltaic load, and historical averages are used to correct for electric vehicle load. Finally, the results of each sub-forecast are linearly superimposed to obtain the composite net load forecast of the virtual power plant. To compensate for the shortcomings of this approach in modeling inter-resource correlations, some schemes introduce deep learning frameworks that fuse multi-source features. These frameworks concatenate the historical load sequences of each sub-resource and feed them into a Transformer or LSTM model for end-to-end learning to capture cross-resource temporal dependencies.
[0005] Because the aforementioned multi-source fusion framework essentially performs independent attention weight calculations for each dimension of the composite feature vector at each time point when processing the time series of each sub-load, the weight allocation is based on the relative amplitude level of the feature value at that time point in the global sequence. When a sudden drop in photovoltaic output coincides with the peak of electric vehicle fast charging and the peak of industrial load in the same short time window, the upward trend of each sub-load's upward rate signal, after being incorporated into the composite feature vector, will have its relative weight reduced by the global sequence mean normalization operation. This results in the structural peak pressure signal generated by the simultaneous upward movement of multiple sources at this specific moment being treated as a random disturbance and smoothed and suppressed by the model. Ultimately, this is reflected in the prediction output, where the model has a systematic underestimation bias of the composite net load peak during the concurrent upward movement of these multi-heterogeneous sub-loads. These peak periods precisely correspond to the critical operating conditions where the virtual power plant faces the highest risk of dispatch deviation assessment. Summary of the Invention
[0006] This invention provides a virtual power plant load forecasting method and system to solve the technical problem in the prior art where the composite net load peak forecast has a structural underestimation bias due to the mean-averaging effect caused by the concurrent positive ramp-up period of multiple heterogeneous sub-loads.
[0007] In a first aspect, the present invention provides a virtual power plant load forecasting method, comprising the following steps: S1 classifies and files the distributed resources under the jurisdiction of the virtual power plant, and calculates the instantaneous ramp-up rate of each type of sub-resource within a set sliding time window to construct a ramp-up rate sequence. S2, based on historical archives, extracts the historical peak period of composite net load, calculates the positive trigger threshold of the ramp rate for each type of sub-resource by combining the instantaneous ramp rate, and calculates the concurrent peak pressure score at the predicted target time based on the positive trigger threshold of the ramp rate to determine the concurrent peak pressure activation flag. S3, based on the concurrent peak pressure activation flag, the original load characteristics of various sub-resources are split into a baseline component and a ramp-up excess component, and the ramp-up excess component is weighted and amplified in combination with the concurrent peak pressure score to construct concurrent peak pressure enhancement features and generate multi-dimensional feature vectors. S4 inputs the multidimensional feature vector into the time series prediction model, outputs the composite net load prediction value, and adds a concurrent peak pressure flag to the prediction period corresponding to the concurrent peak pressure activation flag.
[0008] Its effect is that by actively identifying and restoring the multi-source ramp-up phase alignment effect at the feature engineering level, it improves the structural underestimation problem of the virtual power plant composite net load peak prediction, and makes the prediction results have a stronger peak tracking capability during the concurrent positive ramp-up period of multiple heterogeneous sub-loads.
[0009] Furthermore, the instantaneous crawl rate of various sub-resources is calculated within a set sliding time window to construct a crawl rate sequence, including: Obtain the measured net power of each type of sub-resource at the current moment and the measured net power at the start of the sliding time window; Calculate the difference between the measured net power at the current moment and the measured net power at the beginning moment, and use the ratio of the difference to the width of the sliding time window as the instantaneous climb rate; The instantaneous ramp-up rates of various sub-resources at all historical time points are combined to finally construct a ramp-up rate sequence.
[0010] Its effects are as follows: the standardized calculation logic of instantaneous ramp rate under a limited sliding time window is used to obtain a quantitative ramp rate index by dividing the net power difference between the beginning and end of the window by the window duration, and then integrating the entire historical time series data to form a unified ramp rate sequence; this method unifies the measurement standard of load change rate of different types of distributed resources, filters short-term sampling noise interference, and provides time-aligned and numerically comparable basic quantitative data for subsequent synchronous ramp identification and trigger threshold calculation during peak periods, ensuring the consistency and reliability of multi-source resource ramp behavior analysis.
[0011] Furthermore, based on the instantaneous crawl rate, the positive trigger thresholds for crawl rates specific to various sub-resources are calculated, including: Extract the instantaneous ramp-up rate observations of various sub-resources at each moment within the historical peak period of the composite net load; The number of sub-resource categories with positive instantaneous climb rate observations is counted, and the frequency distribution of positive climb rates for each resource category is obtained in all peak period samples. Based on the frequency distribution of positive climb rates for various resources in all peak period samples, a statistical analysis of the positive value distribution pattern is performed. The instantaneous climb rate observation value is extracted as the median of the positive value distribution, and finally, the positive trigger threshold for climb rates specific to various sub-resources is calculated.
[0012] Its effect is that by deriving an adaptive trigger threshold based on historical peak statistics, the peak pressure discrimination of various resources maintains a symmetrical balance between the false alarm rate and the false alarm rate, thus avoiding a large number of background fluctuations being mistakenly identified as concurrent peak pressure activation.
[0013] Furthermore, based on the positive trigger threshold of the ramp rate, the concurrent peak voltage score at the predicted target time is calculated to determine the concurrent peak voltage activation flag, including: Determine whether the instantaneous ramp rate of each type of sub-resource at the predicted target time is greater than or equal to the corresponding ramp rate positive trigger threshold; The number of sub-resource categories that meet the condition of being greater than or equal to the positive trigger threshold of the climb rate is counted, and the number of sub-resource categories is used as the concurrent peak pressure score; Compare the concurrent peak voltage score with the set activation threshold. If the concurrent peak voltage score is greater than or equal to the activation threshold, the concurrent peak voltage activation flag is set to the activated state, and the concurrent peak voltage activation flag is finally determined.
[0014] Furthermore, the original load characteristics of various sub-resources are decomposed into a baseline component and a ramp-up excess component, including: The average value of each type of sub-resource at the corresponding time of the same day in the historical archives is calculated, and the average value is used as the benchmark component. Calculate the difference between the original load characteristics of each type of sub-resource and the baseline component; The difference is used as the ramp-up excess component, which ultimately completes the decomposition of the original load characteristics.
[0015] Furthermore, the ramp-up excess component is weighted and amplified by combining concurrent peak pressure scores to construct concurrent peak pressure enhancement features and generate multi-dimensional feature vectors, including: When the concurrent peak pressure activation flag is active and the ramp-up excess component is greater than 0, the ratio of the concurrent peak pressure score to the total number of sub-resource categories is calculated as the amplification factor increment. The ramp excess component is amplified by multiplication based on the amplification factor increment, and the amplified ramp excess component is added to the baseline component to obtain the concurrent peak pressure enhancement characteristics. The concurrent peak pressure enhancement features, instantaneous ramp-up rate, concurrent peak pressure score, and scheduling period attributes of various sub-resources are concatenated to generate a multi-dimensional feature vector.
[0016] Its effect is that by enhancing the peak pressure amplitude information retained in the feature vector, the rising signal suppressed by the mean normalization operation during the concurrent peak pressure period is restored, which helps the downstream prediction model to establish an effective recognition capability of peak-in-peak time series patterns during the training phase.
[0017] Furthermore, the multidimensional feature vector is input into the time series prediction model, which outputs a composite net load prediction value. This process also includes training the time series prediction model. Obtain the multidimensional feature vector sequence and the corresponding true composite net load value from the historical training samples; The penalty weight of historical training samples in the loss function is dynamically determined based on the ratio of concurrent peak pressure score to the total number of sub-resource categories. The time series prediction model is iteratively optimized using a loss function, and finally the trained time series prediction model is obtained.
[0018] Its effect is that by combining dynamic loss weights to enhance the learning of high-concurrency peak pressure samples during the training phase, the model prioritizes the prediction accuracy of high-concurrency peak pressure periods during the training convergence process, thereby improving the accuracy of composite net load prediction.
[0019] Furthermore, the distributed resources under the jurisdiction of the virtual power plant are classified and documented, including: Based on differences in response characteristics, the distributed resources under the jurisdiction of the virtual power plant are divided into random disturbance resources, behavior-driven resources, and planned scheduling resources. Establish separate historical load data archives for each type of resource after the division; Align the time resolution of the historical load data archives to the set scheduling declaration granularity, and finally complete the classification and archiving.
[0020] Furthermore, a concurrent peak voltage flag is added to the prediction period corresponding to the concurrent peak voltage activation flag, including: The composite net load forecast values are time-aligned according to the set scheduling declaration granularity; Filter out the target prediction period when the concurrent peak pressure activation flag is in an active state; When submitting the plan before the submission date, concurrent peak pressure markers are generated and attached for the target forecast period.
[0021] Secondly, the present invention provides a virtual power plant load forecasting system, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned virtual power plant load forecasting method is implemented.
[0022] The beneficial effects are as follows: By adding concurrent peak voltage markers to the predicted output, this invention provides a reference for reserving reserve capacity in the dispatching system, thereby reducing the risk of dispatch deviation assessment for virtual power plants under high-risk operating conditions. Simultaneously, through multi-dimensional feature extraction and dynamic weight adjustment, this invention enables the prediction model to accurately capture the complex temporal characteristics of concurrent positive ramp-up of diverse heterogeneous sub-loads, improving the overall accuracy and stability of load forecasting. Attached Figure Description
[0023] Figure 1 This is a flowchart of a virtual power plant load forecasting method.
[0024] Figure 2 This is a comparison curve of the combined net load prediction of existing methods and the method of this invention. Detailed Implementation
[0025] 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, not all, of the embodiments of the present invention. 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.
[0026] An embodiment of the virtual power plant load forecasting method provided by this invention: like Figure 1 As shown, the virtual power plant load forecasting method includes the following steps: S1, classify and archive resources and construct a crawl rate sequence.
[0027] The distributed resources under the jurisdiction of the virtual power plant are classified into three categories based on their response characteristics: random disturbance resources, behavior-driven resources, and planned dispatch resources. Random disturbance resources include distributed photovoltaic and wind turbines, whose output is directly driven by weather conditions, exhibiting a large ramp-up slope and short duration. Behavior-driven resources include electric vehicle charging piles and building HVAC systems, whose load changes are triggered by user behavior and temperature signals, exhibiting a concentrated time clustering pattern. Planned dispatch resources include industrial flexible loads and energy storage power stations, whose ramp-up characteristics are constrained by dispatch plans, resulting in relatively regular changes. Independent historical load data archives are maintained for each of these three types of resources, and the time resolution is aligned to a set dispatch reporting granularity. For example, the dispatch reporting granularity can be set to 15 minutes to maintain consistency with the grid dispatch reporting cycle, ultimately completing the classification and archiving process.
[0028] For each type of sub-resource, calculate its instantaneous ramp rate within a set sliding time window. Obtain the measured net power of each type of sub-resource at the current moment and the measured net power at the start of the sliding time window. Taking sub-resources as an example, in the case of The time is the endpoint, and the width is Within the time window, calculate the difference between the measured net power at the current moment and the measured net power at the beginning moment, and use the ratio of this difference to the width of the sliding time window as the instantaneous climb rate. Defined as:
[0029] in, for The measured net power of this type of sub-resource at any given time, with power generation being negative and power consumption being positive, and the unit is MW. This represents the measured net power at the start of the sliding time window. The granularity of the scheduling request is determined by taking twice the granularity of the scheduling request. For example, when the granularity of the scheduling request is 15 minutes... A 30-minute interval is acceptable; the difference between two consecutive reporting granularities can fully cover the ramp-up actions within a single reporting cycle and effectively filter out second-level spikes caused by sampling noise. If the reporting cycle of the power grid connected to the virtual power plant changes, then... The particle size was subsequently adjusted to twice that of the newly submitted application.
[0030] A positive value indicates that the net electricity demand for this type of sub-resource continues to increase within the window, i.e., net load ramp-up; A negative value indicates a continuous increase in net supply, i.e., a decrease in net load. For each type of sub-resource, the instantaneous ramp-up rate of each sub-resource at all historical time points is calculated. By combining these data, a climb rate sequence is ultimately constructed. and the original load sequence These resources are jointly stored in the database for this type of sub-resource, serving as the basic input for subsequent peak concurrency identification. It should be noted that those skilled in the art can reasonably adjust the classification method based on the actual types of resources accessed by the virtual power plant; this embodiment does not limit the specific number of resource categories.
[0031] By independently archiving the historical load data of various sub-resources and calculating the ramp rate sequence within a fixed-width time window, a quantitative basis input with a unified time resolution is provided for subsequent identification of multi-source co-current ramp events.
[0032] S2, calculate the trigger threshold and score to determine the activation flag.
[0033] Historical peak periods of the composite net load are extracted from historical archives as a reference sample set. The peak period is defined as the set of times in history when the composite net load exceeds its annual average plus a preset standard deviation. This standard deviation is calculated from all historical net load data; for example, the preset multiple can be 1.5. Combined with the instantaneous climb rate, positive trigger thresholds for climb rates specific to various sub-resources are calculated. Specifically, this involves extracting every moment falling within the historical peak periods of the composite net load. Instantaneous ramp-up rate observations for various sub-resources The number of sub-resource categories with positive instantaneous climb rate observations is counted, and the frequency distribution of positive climb rates for each resource category is obtained in all peak period samples, serving as a statistical benchmark for historical peak characteristics. For each sub-resource category... Based on the frequency distribution mentioned above, a statistical analysis of the positive value distribution pattern is performed. The instantaneous climb rate observation value is extracted as the median of the positive value distribution. Finally, the positive trigger threshold for climb rate specific to each type of sub-resource is calculated. .
[0034] Calculate the predicted target time based on the positive trigger threshold of the climb rate. The concurrent peak pressure score is used to determine the concurrent peak pressure activation flag. It is determined whether the instantaneous ramp rate of each type of sub-resource at the predicted target time is greater than or equal to the corresponding positive ramp rate trigger threshold. The number of sub-resource categories that meet the condition of being greater than or equal to the positive ramp rate trigger threshold is counted, and this number of sub-resource categories is used as the concurrent peak pressure score. :
[0035] in, The total number of sub-resource categories, This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Representative at The number of sub-resource categories that simultaneously surpass their respective historical peak climbing median thresholds, with a value range of [value range missing]. A larger value indicates that more sub-resource categories are experiencing synchronous positive ramp-ups at the current moment, resulting in stronger concurrent peak pressure on the composite net load; a smaller value indicates that the ramp-ups of each sub-resource are more dispersed, resulting in a weaker peak superposition effect of the composite net load. achieve When this occurs, it indicates that all categories of sub-resources are simultaneously climbing upwards, representing the most severe state of peak concurrent pressure.
[0036] Further compare the concurrent peak voltage score with the set activation threshold. If the concurrent peak voltage score is greater than or equal to the activation threshold, then set the concurrent peak voltage activation flag to the active state, thus finally determining the concurrent peak voltage activation flag. : in response to When the preset activation threshold is reached, This indicates that the system is currently in a state of peak concurrent pressure activation; otherwise... For example, an activation threshold is desirable. This means that activation is triggered when more than half of the categories are climbing synchronously.
[0037] By deriving specific trigger thresholds for various sub-resources based on historical peak samples, and using the count of categories exceeding the threshold as the concurrent peak pressure score, effective identification of multi-source in-phase climbing events is achieved.
[0038] S3, split the load characteristics and construct concurrent peak voltage enhancement characteristics.
[0039] Based on the concurrent peak pressure activation flag, the original load characteristics of various sub-resources are split into a baseline component and a ramp-up excess component. At that time, the load characteristics of various sub-resources are fed into the prediction model according to the standard normalization method; when At that time, positive weight amplification is applied to the climbing components that exceed the trigger threshold for various sub-resources. For the first... For each type of sub-resource, the average value of the corresponding time on the same day in the historical archives is calculated, and this average value is used as the baseline component. Weekdays and holidays are counted separately. When the number of days of the same type in the historical archives is insufficient for the preset reference window, all available samples are used; when it exceeds the preset reference window, samples from the most recent few days are used to ensure that the benchmark component tracks recent changes in energy consumption patterns. For example, the preset reference window can be 20 days. The original load characteristics of various sub-resources are calculated. The difference between the baseline component and the baseline component will be used as the climb excess component. Finally, the original load characteristics are decomposed and defined as follows:
[0040] By combining concurrent peak pressure scores with weighted amplification of the ramp excess component, concurrent peak pressure enhancement features are constructed and multi-dimensional feature vectors are generated. When and At that time, the ratio of concurrent peak pressure score to the total number of sub-resource categories is calculated as the amplification factor increment. Based on the amplification factor increment, the ramp excess component is multiplied and amplified. The amplified ramp excess component is then added to the baseline component to obtain the concurrent peak pressure enhancement feature. :
[0041] in, The proportion of concurrently activated sub-resource categories, with a value range of [value range missing]. The contribution of a single-category excess ramp component to the peak composite net load is positively correlated with the number of categories participating in concurrent ramps. This can be further explained by introducing... As an amplification factor increment, the climb excess component is corrected upwards.
[0042] when and If, during peak concurrency periods, this type of sub-resource actually shows a net increase in supply, then its original characteristics are maintained. Keep it unchanged to avoid compressing beneficial peak-shaving components. When At that time, maintain for all categories .
[0043] Enhance the concurrent peak voltage characteristics of various sub-resources Instantaneous climb rate Concurrent peak pressure score The scheduling period attributes are concatenated to generate a multi-dimensional feature vector, which is then input into the subsequent prediction model.
[0044] By scaling up the excess component of the concurrent peak pressure activation period at the feature engineering level, the prediction model can sense the peak pressure amplitude of multi-source in-phase climbing events when receiving the enhanced feature vector, thus restoring the climbing signal suppressed by the mean normalization operation during the concurrent peak pressure period.
[0045] S4 outputs the composite net load forecast and adds a label.
[0046] Before inputting the multidimensional feature vectors into the time-series prediction model, the model is first trained: The multidimensional feature vector sequence and corresponding true composite net load values from historical training samples are obtained. The penalty weight of historical training samples in the loss function is dynamically determined based on the ratio of concurrent peak pressure scores to the total number of sub-resource categories. The loss function is then used to iteratively optimize the time-series prediction model, ultimately resulting in a trained model. Since the ramp-up excess component during concurrent peak pressure activation periods has been amplified and restored at the feature level, the prediction model can identify the composite net load peak shape corresponding to the synchronous over-threshold ramp-up rate of multiple sub-resources in historical training samples. This allows for the prediction of corresponding peak amplitudes for future periods with similar concurrent activation characteristics during the prediction phase. During model training, the loss function applies penalty weights to the concurrent peak pressure activation samples. These weights are determined by… Determined dynamically. The expected peak deviation of the composite net load during synchronous ramp-up of all categories is approximately equal to the deviation of a single category. The multiple corresponds to the loss contribution and the concurrency ratio. They are positively correlated, therefore... This serves as the increment of the loss weight for that sample. The higher the concurrency of historical training samples, the greater the contribution weight of their prediction error in the loss function, ensuring that the model prioritizes the prediction accuracy during periods of high concurrency peaks during training convergence.
[0047] The enhanced feature vector sequence is input into the trained time-series prediction model, which outputs the composite net load forecast for the next reporting period one day in advance. It should be noted that there are many other models in this field capable of performing sequence prediction. Those skilled in the art can choose different sequence prediction models according to their actual needs. Time series Transformer, LSTM, etc., can all be used as optional implementation methods for time series prediction models.
[0048] The concurrent peak pressure flag is attached to the prediction period corresponding to the concurrent peak pressure activation flag. Specifically, this involves: aligning the composite net load forecast value to the set scheduling declaration granularity; selecting the target prediction period where the concurrent peak pressure activation flag is active; and generating and attaching the concurrent peak pressure flag to the target prediction period when submitting the plan before submission, for the scheduling system to reference when formulating reserve capacity reservation plans. It should be noted that those skilled in the art can reasonably adjust the attachment form and transmission method of the concurrent peak pressure flag according to the interface specifications of the actual scheduling system.
[0049] Thus, by introducing the concurrent peak pressure enhancement feature into the training and inference process of the prediction model, and combining dynamic loss weights to strengthen the learning of high-concurrency peak pressure samples during the training phase, the peak tracking capability of composite net load prediction during the concurrent positive ramp-up period of multi-heterogeneous sub-loads is improved.
[0050] like Figure 2 As shown, during normal periods when there is no simultaneous ramp-up of multiple resources, the prediction results of both existing methods and the method of this invention are consistent with the actual load. During periods of concurrent peak load activation, the simultaneous ramp-up of multiple types of distributed resources forms load spikes. Existing methods have significant peak underestimation due to the normalization suppression of ramp-up signals. This invention enhances the peak load characteristics, making the prediction curve highly consistent with the actual composite net load, effectively eliminating peak prediction bias, and marking high-risk peak intervals, providing data support for reserving backup capacity for virtual power plant dispatch.
[0051] An embodiment of the virtual power plant load forecasting system provided by this invention: The virtual power plant load forecasting system includes a processor and a memory. The memory stores computer program instructions, which are executed by the processor to implement the virtual power plant load forecasting method described above.
[0052] The virtual power plant load forecasting system also includes other components well known to those skilled in the art, such as communication interfaces. Their setup and functions are known in the art and will not be described in detail here.
[0053] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A virtual power plant load forecasting method, characterized in that, Includes the following steps: S1 classifies and files the distributed resources under the jurisdiction of the virtual power plant, and calculates the instantaneous ramp-up rate of each type of sub-resource within a set sliding time window to construct a ramp-up rate sequence. S2, based on historical archives, extract the historical peak period of composite net load, combine the instantaneous ramp rate to calculate the ramp rate positive trigger threshold for each type of sub-resource, and calculate the concurrent peak pressure score at the predicted target time based on the ramp rate positive trigger threshold to determine the concurrent peak pressure activation flag, including: judging whether the instantaneous ramp rate of each type of sub-resource at the predicted target time is greater than or equal to the corresponding ramp rate positive trigger threshold; The number of sub-resource categories that meet the condition of being greater than or equal to the positive trigger threshold of the climb rate is counted, and the number of sub-resource categories is used as the concurrent peak pressure score; Compare the concurrent peak voltage score with the set activation threshold. If the concurrent peak voltage score is greater than or equal to the activation threshold, the concurrent peak voltage activation flag is set to the activated state, and the concurrent peak voltage activation flag is finally determined. S3, based on the concurrent peak pressure activation flag, decomposes the original load characteristics of various sub-resources into a baseline component and a ramp-up excess component, and combines the concurrent peak pressure score to apply weight amplification to the ramp-up excess component, in order to construct concurrent peak pressure enhancement features and generate a multi-dimensional feature vector, including: When the concurrent peak pressure activation flag is active and the ramp-up excess component is greater than 0, the ratio of the concurrent peak pressure score to the total number of sub-resource categories is calculated as the amplification factor increment. The ramp excess component is amplified by multiplication based on the amplification factor increment, and the amplified ramp excess component is added to the baseline component to obtain the concurrent peak pressure enhancement characteristics. The concurrent peak pressure enhancement features, instantaneous ramp-up rate, concurrent peak pressure score, and scheduling period attributes of various sub-resources are concatenated to finally generate a multi-dimensional feature vector; S4 inputs the multidimensional feature vector into the time series prediction model, outputs the composite net load prediction value, and adds a concurrent peak pressure flag to the prediction period corresponding to the concurrent peak pressure activation flag.
2. The virtual power plant load forecasting method according to claim 1, characterized in that, Calculate the instantaneous crawl rate of various sub-resources within a set sliding time window to construct a crawl rate sequence, including: Obtain the measured net power of each type of sub-resource at the current moment and the measured net power at the start of the sliding time window; Calculate the difference between the measured net power at the current moment and the measured net power at the beginning moment, and use the ratio of the difference to the width of the sliding time window as the instantaneous climb rate; The instantaneous ramp-up rates of various sub-resources at all historical time points are combined to finally construct a ramp-up rate sequence.
3. The virtual power plant load forecasting method according to claim 1, characterized in that, Calculate the positive trigger threshold for crawl rate specific to various sub-resources based on the instantaneous crawl rate, including: Extract the instantaneous ramp-up rate observations of various sub-resources at each moment within the historical peak period of the composite net load; The number of sub-resource categories with positive instantaneous climb rate observations is counted, and the frequency distribution of positive climb rates for each resource category is obtained in all peak period samples. Based on the frequency distribution of positive climb rates for various resources in all peak period samples, a statistical analysis of the positive value distribution pattern is performed. The instantaneous climb rate observation value is extracted as the median of the positive value distribution, and finally, the positive trigger threshold for climb rates specific to various sub-resources is calculated.
4. The virtual power plant load forecasting method according to claim 1, characterized in that, The original load characteristics of various sub-resources are broken down into a baseline component and a ramp-up excess component, including: The average value of each type of sub-resource at the corresponding time of the same day in the historical archives is calculated, and the average value is used as the benchmark component. Calculate the difference between the original load characteristics of each type of sub-resource and the baseline component; The difference is used as the ramp-up excess component, which ultimately completes the decomposition of the original load characteristics.
5. The virtual power plant load forecasting method according to claim 1, characterized in that, The multidimensional feature vector is input into the time series prediction model, which outputs a composite net load prediction value. This process also includes training the time series prediction model. Obtain the multidimensional feature vector sequence and the corresponding true composite net load value from the historical training samples; The penalty weight of historical training samples in the loss function is dynamically determined based on the ratio of concurrent peak pressure score to the total number of sub-resource categories. The time series prediction model is iteratively optimized using a loss function, and finally the trained time series prediction model is obtained.
6. The virtual power plant load forecasting method according to claim 1, characterized in that, The distributed resources under the jurisdiction of the virtual power plant are classified and documented, including: Based on differences in response characteristics, the distributed resources under the jurisdiction of the virtual power plant are divided into random disturbance resources, behavior-driven resources, and planned scheduling resources. Establish separate historical load data archives for each type of resource after the division; Align the time resolution of the historical load data archives to the set scheduling declaration granularity, and finally complete the classification and archiving.
7. The virtual power plant load forecasting method according to claim 1, characterized in that, Add concurrent peak voltage markers to the prediction period corresponding to the concurrent peak voltage activation flag, including: The composite net load forecast values are time-aligned according to the set scheduling declaration granularity; Filter out the target prediction period when the concurrent peak pressure activation flag is in an active state; When submitting the plan before the submission date, concurrent peak pressure markers are generated and attached for the target forecast period.
8. A virtual power plant load forecasting system, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the virtual power plant load forecasting method according to any one of claims 1-7 is implemented.
Citation Information
Patent Citations
Virtual power plant-oriented multi-data source fusion multi-energy load prediction method
CN121958817A
Virtual power plant resource aggregation method for dynamic peak regulation demand of power grid
CN121212717A
Multi-element heterogeneous resource aggregation method and device for energy storage type virtual power plant
CN121581382A
