Flexible resource benchmark electric quantity probability prediction method and device, computer device, medium and product

Abstract

The application relates to a flexible resource benchmark power probability prediction method and device, computer equipment, a medium and a product. The method comprises the following steps: obtaining to-be-predicted data of a to-be-predicted flexible resource type according to the to-be-predicted flexible resource type, that is, historical operation data of the to-be-predicted flexible resource and associated influence factor data; processing the to-be-predicted data based on normal quantile transformation to obtain target prediction data; extracting multiple groups of model parameters from a model parameter posterior distribution of a pre-trained benchmark power probability prediction model, and performing a complete forward propagation on each group of model parameters in combination with the target prediction data to obtain a corresponding prediction Gaussian distribution; generating a corresponding prediction initial result according to each prediction Gaussian distribution, and processing each prediction initial result based on inverse normal quantile transformation to obtain a benchmark power probability prediction result of the to-be-predicted flexible resource. The method can accurately predict the benchmark power of multiple types of flexible resources.

Description

Technical Field

[0001] This application relates to the field of baseline energy probability prediction technology, and in particular to a flexible resource baseline energy probability prediction method, apparatus, computer equipment, medium and product. Background Technology

[0002] With the large-scale grid connection of flexible resources such as distributed photovoltaics, electric vehicle charging piles, and smart building central air conditioning, accurate prediction of the benchmark electricity of flexible resources has become an important technical means for grid dispatch and demand response.

[0003] In traditional technologies, in order to achieve accurate prediction of the baseline electricity of flexible resources, the historical operating data of the power grid is usually preprocessed and then directly input into the Bayesian prediction model for training. The probability prediction is completed using a unified model structure and parameter configuration, and the prediction results and corresponding probability intervals are obtained based on the model output.

[0004] However, traditional methods for predicting baseline electricity levels for flexible resources treat all types of resources uniformly, resulting in low prediction accuracy. Therefore, there is an urgent need to propose a probabilistic prediction method for baseline electricity levels for multiple types of flexible resources. Summary of the Invention

[0005] Therefore, it is necessary to provide a flexible resource baseline power probability prediction method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can perform baseline power probability prediction for multiple types of flexible resources, addressing the aforementioned technical problems.

[0006] In a first aspect, this application provides a method for predicting the probability of baseline electricity consumption for flexible resources, comprising: obtaining prediction data for the flexible resource to be predicted based on its resource type; the prediction data including historical operating data and associated influencing factor data of the flexible resource to be predicted; processing the prediction data based on normal quantile transformation to obtain target prediction data; extracting multiple sets of model parameters from the posterior distribution of model parameters of a pre-trained baseline electricity consumption probability prediction model, and performing a complete forward propagation on each set of model parameters in conjunction with the target prediction data to obtain prediction Gaussian distributions corresponding to each set of model parameters; generating corresponding initial prediction results based on each prediction Gaussian distribution, and processing each initial prediction result based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline electricity consumption probability prediction result for the flexible resource to be predicted.

[0007] In one embodiment, the step of processing each of the initial prediction results based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline energy probability prediction result of the flexible resource to be predicted includes: processing each of the initial prediction results based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline energy probability prediction result to be constrained for the flexible resource to be predicted; and performing physical constraint recalibration on the baseline energy probability prediction result of the flexible resource to be predicted according to the preset physical constraints corresponding to the flexible resource to be predicted to obtain the baseline energy probability prediction result of the flexible resource to be predicted.

[0008] In one embodiment, the step of recalibrating the prediction result to be constrained based on the preset physical constraints corresponding to the flexible resource to be predicted, to obtain a baseline power probability prediction result for the flexible resource to be predicted, includes: determining the corresponding preset physical constraints based on the resource type of the flexible resource to be predicted; truncating and correcting the prediction result to be constrained based on the preset physical constraints to obtain a truncated intermediate prediction result; evaluating the truncated coverage loss of the intermediate prediction result to obtain a truncated coverage loss evaluation result; and performing adaptive interval width compensation on the intermediate prediction result based on the truncated coverage loss evaluation result to obtain a baseline power probability prediction result for the flexible resource to be predicted.

[0009] In one embodiment, the pre-trained baseline power probability prediction model is trained as follows: Training data for each sample flexible resource is acquired, including historical operational sample data and associated influencing factor sample data for the sample flexible resource; a normal quantile transformation is performed on each training data to obtain corresponding target training data; prior parameters of the model parameters of the baseline power probability prediction model are determined according to the resource type of each sample flexible resource, and a prior distribution of the model parameters is established based on the prior parameters; the baseline power probability prediction model is trained based on the target training data and the prior distribution of the model parameters to construct a variational posterior distribution of the model parameters; variational inference is used to learn the variational posterior distribution of the model parameters, and the model parameters are iteratively updated until a preset training termination condition is met, resulting in a trained baseline power probability prediction model.

[0010] In one embodiment, determining the prior parameters of the model parameters of the baseline power probability prediction model to be trained based on the resource type of each of the sample flexible resources, and establishing the prior distribution of the model parameters based on the prior parameters, includes: determining the peak-thick plate prior parameters of the model parameters of the baseline power probability prediction model to be trained based on the resource type of each of the sample flexible resources, and establishing the prior distribution of the model parameters based on the peak-thick plate prior parameters.

[0011] In one embodiment, training the baseline battery probability prediction model based on the target training data and the prior distribution of the model parameters includes: inputting the target training data into the baseline battery probability prediction model, and training the target training data sequentially based on the input layer, one-dimensional convolutional layer, max pooling layer, bidirectional long short-term memory network layer, and dense output layer of the baseline battery probability prediction model, combined with the prior distribution of the model parameters; wherein: the input layer is used to receive the target training data and input the target training data into the one-dimensional convolutional layer; the one-dimensional convolutional layer is used to extract local features from the target training data; the max pooling layer is used to perform feature compression processing on the extracted results obtained from the local feature extraction; the bidirectional long short-term memory network layer is used to learn temporal features from the compressed features after feature compression processing; and the dense output layer is used to output predicted Gaussian distribution parameters.

[0012] Secondly, this application also provides a flexible resource-based energy probability prediction device, comprising:

[0013] The data acquisition module is used to acquire the forecast data of the flexible resource to be predicted based on the resource type of the flexible resource to be predicted; the forecast data includes the historical operation data of the flexible resource to be predicted and the associated influencing factor data.

[0014] The quantile transformation module is used to process the data to be predicted based on the normal quantile transformation to obtain the target prediction data;

[0015] The probability prediction module is used to extract multiple sets of model parameters from the posterior distribution of model parameters of the pre-trained benchmark power probability prediction model, and perform a complete forward propagation on each set of model parameters in combination with the target prediction data to obtain the prediction Gaussian distribution corresponding to each set of model parameters.

[0016] The result generation module is used to generate corresponding initial prediction results based on each of the predicted Gaussian distributions, and to process each of the initial prediction results based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline power probability prediction result of the flexible resource to be predicted.

[0017] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method steps of any one of the first aspects.

[0018] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method steps of any one of the first aspects.

[0019] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the method steps of any one of the first aspects.

[0020] The aforementioned method, apparatus, computer equipment, computer-readable storage medium, and computer program product for predicting the baseline electricity consumption of flexible resources obtain the data to be predicted for the flexible resources by identifying the resource type of the flexible resources to be predicted, which helps to improve the correlation between the acquired data to be predicted and the features of the flexible resources. Preprocessing the data using normal quantile transformation aligns the data distribution with the prior distribution, improving the data adaptability of the pre-trained baseline electricity consumption probability prediction model. By extracting multiple sets of model parameters based on the posterior distribution of the model parameters from the pre-trained baseline electricity consumption probability prediction model, and performing forward propagation on each set of model parameters using the target prediction data, the corresponding prediction Gaussian distribution is obtained. This represents the uncertainty of the model parameters and provides initial probability prediction of the baseline electricity consumption, effectively reflecting the reliability of the prediction results. Furthermore, by using inverse normal quantile transformation to restore the initial prediction results corresponding to the initial probability prediction to the original data space, the obtained baseline electricity consumption probability prediction results are more consistent with the actual operating data distribution, thereby improving the accuracy and reliability of the baseline electricity consumption prediction for flexible resources. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is an application environment diagram of a flexible resource-based energy probability prediction method in one embodiment;

[0023] Figure 2 This is a flowchart illustrating a flexible resource-based energy probability prediction method in one embodiment.

[0024] Figure 3 This is a schematic diagram of the model training process for a baseline electricity probability prediction model in one embodiment.

[0025] Figure 4 This is a schematic diagram of the structure of a Bayesian convolutional neural network and a bidirectional long short-term memory network model in one embodiment;

[0026] Figure 5 This is a structural block diagram of a flexible resource-based energy probability prediction device in one embodiment;

[0027] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0030] Distributed flexible resources in practical applications exhibit significant heterogeneity, with substantial differences in power variation patterns and influencing factors across different resource types. Current benchmark power forecasting methods for flexible resources typically employ uniform data features and model structures to predict benchmark power for different types of flexible resources, making it difficult to fully characterize the operational characteristics of various resources. Furthermore, the actual collected resource operation data, meteorological data, and user behavior data may contain missing data, noise, and anomalies, further increasing the uncertainty of benchmark power forecasting. Existing methods often treat data preprocessing, probabilistic modeling, and physical constraints as independent steps, lacking a unified and collaborative design. Under these circumstances, a single point forecast result is insufficient to effectively characterize the forecast risk of benchmark power. Although some probabilistic forecasting methods can output a forecast distribution or probability interval, if the physical operational constraints of different resource types are not fully considered, the obtained probabilistic forecast results may exceed the actual feasible operational range, thereby affecting the reliability of the forecast results and the scheduling and demand response based on the forecast results.

[0031] This application provides a flexible resource-based energy probability prediction method, which can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. The data storage system can store the data that server 104 needs to process, such as historical operating data of various flexible resources and related influencing factor data. The data storage system can be integrated on server 104 or placed on the cloud or other network servers. Server 104 is used to acquire the unpredictable data of the flexible resources to be predicted collected by terminal 102; the unpredictable data includes historical operating data of the flexible resources to be predicted and related influencing factor data; it processes the unpredictable data based on normal quantile transformation to obtain target prediction data; it extracts multiple sets of model parameters from the posterior distribution of the model parameters of the pre-trained baseline power probability prediction model, and performs a complete forward propagation on each set of model parameters in conjunction with the target prediction data to obtain the prediction Gaussian distribution corresponding to each set of model parameters; it generates corresponding initial prediction results based on each prediction Gaussian distribution, and processes each initial prediction result based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline power probability prediction result of the flexible resources to be predicted. Terminal 102 can be, but is not limited to, various personal computers, laptops, and tablets. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.

[0032] In one exemplary embodiment, such as Figure 2 As shown, a flexible resource-based baseline energy probability prediction method is provided, which can be applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 202 to 208. Wherein:

[0033] Step 202: Based on the resource type of the flexible resource to be predicted, obtain the data to be predicted for the flexible resource to be predicted; the data to be predicted includes the historical operating data of the flexible resource to be predicted and the related influencing factor data.

[0034] Flexible resources include central air conditioning, electric vehicle charging stations, photovoltaic power generation, etc.; resource type is used to indicate the category to which the flexible resource belongs. Different types of flexible resources have different operating rules and environmental impact factors, and the data types of the associated impact factors are also different.

[0035] For example, since the baseline influencing factors and operating patterns of different types of resources differ significantly, the server needs to obtain historical operating data and associated influencing factor data of the target flexible resource (the flexible resource to be predicted) according to the resource type:

[0036] Photovoltaic resources: Historical operating data includes photovoltaic inverter power data; related influencing factor data includes meteorological characteristics such as irradiance, temperature, humidity, and wind speed, as well as time characteristics such as month and hour;

[0037] Electric vehicle charging station resources: Historical operation data includes the total power of the charging station and the charging status data of each charging pile; related influencing factor data includes meteorological characteristics such as temperature and rainfall, as well as time characteristics such as holidays, days of the week, and hours;

[0038] Central air conditioning resources: Historical operating data includes the start-stop status and power data of air conditioning equipment; related influencing factor data includes meteorological characteristics such as outdoor temperature and humidity, as well as time characteristics such as month and hour. The air conditioning load has strong temperature-power coupling and continuous fluctuation.

[0039] Step 204: Process the data to be predicted based on normal quantile transformation to obtain the target prediction data.

[0040] Among them, the Normal Quantile Transform (NQT) is a method for mapping data distribution. This method first calculates the quantile value of the original data in the empirical cumulative distribution, and then uses the inverse cumulative distribution function of the standard normal distribution to map the original data to a data space that approximately follows the standard normal distribution.

[0041] For example, after the server-acquired data to be predicted undergoes a preprocessing process (including outlier handling, missing value imputation, and standardization), the NQT method is used to perform distribution mapping on the preprocessed data. First, the cumulative probability corresponding to each data sample is obtained through the Empirical Cumulative Distribution Function (ECDF). Then, the inverse cumulative distribution function of the standard normal distribution is used to map various data samples to the standard normal distribution space, making the transformed data approximately conform to the standard normal distribution, thereby obtaining the target prediction data. Specifically, the preprocessing process includes: the server first identifies and removes outliers from the data to be predicted, imputs missing data using linear interpolation, and then standardizes each variable to satisfy a distribution with zero mean and unit variance, thereby eliminating the influence of different units on model training.

[0042] Step 206: Extract multiple sets of model parameters from the posterior distribution of model parameters of the pre-trained baseline power probability prediction model, and perform a complete forward propagation on each set of model parameters in conjunction with the target prediction data to obtain the prediction Gaussian distribution corresponding to each set of model parameters.

[0043] Among them, the baseline electricity probability prediction model is a pre-trained model; the posterior distribution of model parameters is the probability distribution of model parameters obtained by variational inference during the model training phase, which is used to represent the uncertainty of model parameters.

[0044] For example, in the prediction phase of the baseline electricity probability prediction model, the server randomly samples multiple sets of model parameters from the posterior distribution of model parameters obtained during the model training phase, with each set of model parameters corresponding to a neural network instance. Subsequently, each set of sampled model parameters is input into the model to perform a complete forward propagation. Each forward propagation outputs a set of predicted Gaussian distribution parameters, i.e., predicted mean parameters. and prediction variance parameter This forms a predictive Gaussian distribution N( , After multiple samplings, multiple predicted Gaussian distributions can be obtained. These Gaussian distributions simultaneously represent the cognitive uncertainty caused by the change in the predicted mean introduced by the posterior sampling of model parameters and the random uncertainty brought about by the quantification of the output distribution variance.

[0045] Step 208: Generate corresponding initial prediction results based on each predicted Gaussian distribution, and process each initial prediction result based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline power probability prediction result of the flexible resource to be predicted.

[0046] The Inverse Normal Quantile Transformation (InverseNQT) is the inverse process of the normal quantile transformation, used to restore the prediction results in the standard normal space to the original data space.

[0047] For example, the server performs an inverse normal quantile transformation on each initial prediction result to obtain the corresponding original data space prediction result; calculates the prediction mean based on each original data space prediction result, and uses the prediction mean as the baseline power point prediction result for the flexible resource to be predicted; performs statistical analysis on each original data space prediction result, and determines the corresponding prediction interval according to a preset confidence level. For example, at an 85% confidence level, the 7.5% quantile of the original data space prediction result is used as the lower bound of the prediction interval, and the 92.5% quantile is used as the upper bound of the prediction interval, thus obtaining the probabilistic prediction interval for the baseline power of the flexible resource to be predicted. The output baseline power probability prediction result includes: the baseline power point prediction value (truncated sample mean), the prediction mean and prediction variance, the probability prediction interval (lower and upper bounds) at the 85% confidence level after recalibration, and the truncated coverage loss assessment result (…). , , By performing inverse normal quantile transformations on each initial prediction result, the probability prediction interval in the original data space can be kept consistent with the probability coverage in the normal distribution space, avoiding the coverage deviation caused by directly performing inverse transformations on the prediction interval boundaries. When other confidence levels are required, the corresponding probability prediction interval can be determined based on the relevant quantiles.

[0048] In the aforementioned method for probabilistic prediction of benchmark electricity consumption for flexible resources, the model obtains the data to be predicted based on the resource type of the flexible resource to be predicted. This allows the model to comprehensively consider the baseline influencing factors and operational patterns of different resource types, improving the ability of the data to represent the variation patterns of benchmark electricity consumption and providing a reliable data foundation for subsequent predictions. Preprocessing the data to be predicted using normal quantile transformation aligns the data distribution with the prior distribution, improving the model's data adaptability and enhancing the stability of the prediction process. By extracting multiple sets of model parameters based on the posterior distribution of model parameters and performing forward propagation on each, the corresponding predicted Gaussian distribution is obtained. This not only yields a single predicted value but also reflects the credibility of the prediction results, improving their reliability. Initial prediction results are generated based on the predicted Gaussian distribution, and each initial prediction result undergoes inverse normal quantile transformation. This ensures that the coverage of the original domain probability interval is consistent with the normal domain, avoiding coverage deviations that may be caused by inverse quantile transformation. Statistical analysis is then performed on multiple prediction results to obtain point prediction results of the benchmark electricity consumption and the corresponding probability prediction intervals. This more accurately reflects the uncertainty of benchmark electricity consumption prediction and enhances the practical application value of the prediction results.

[0049] In an exemplary embodiment, each initial prediction result is processed based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline energy probability prediction result of the flexible resource to be predicted. This includes: processing each initial prediction result based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline energy probability prediction result of the flexible resource to be predicted; and performing physical constraint recalibration on the baseline energy probability prediction result of the flexible resource to be predicted according to the preset physical constraints corresponding to the flexible resource to be predicted to obtain the baseline energy probability prediction result of the flexible resource to be predicted.

[0050] Among them, preset physical constraints refer to constraints predetermined based on the specific characteristics of different resource types, used to limit the physical feasibility range that the baseline power prediction results should meet. Different types of flexible resources correspond to different physical constraints.

[0051] For example, before obtaining the constrained prediction result of the baseline energy level of the flexible resource to be predicted, the server processes each initial prediction result based on the inverse normal quantile transformation corresponding to the normal quantile transformation, and then performs physical constraint recalibration. First, the server obtains the corresponding physical operating boundary conditions according to the resource type of the flexible resource to be predicted, and performs constraint correction on the part that exceeds the corresponding physical boundary, so that the prediction result after constraint correction meets the actual operating constraints, and finally obtains the baseline energy level probability prediction result of the flexible resource to be predicted.

[0052] In this embodiment, the baseline power is obtained by using a baseline power probability prediction model. Preset physical constraints corresponding to the type of flexible resource to be predicted are introduced, and the physical constraints are recalibrated to ensure that the prediction results meet the physical boundary constraints in the actual operation of the flexible resource.

[0053] In an exemplary embodiment, based on the preset physical constraints corresponding to the flexible resource to be predicted, the physical constraint recalibration of the constraint prediction result is performed to obtain the baseline power probability prediction result of the flexible resource to be predicted. This includes: determining the corresponding preset physical constraints based on the resource type of the flexible resource to be predicted; truncating and correcting the constraint prediction result based on the preset physical constraints to obtain the truncated intermediate prediction result; evaluating the truncated coverage loss of the intermediate prediction result to obtain the truncated coverage loss evaluation result; and performing adaptive interval width compensation on the intermediate prediction result based on the truncated coverage loss evaluation result to obtain the baseline power probability prediction result of the flexible resource to be predicted.

[0054] Among them, truncation correction refers to the process of adjusting the predicted values ​​that exceed the preset physical constraint range to the corresponding physical boundary range when some predicted values ​​in the baseline electricity probability prediction result exceed the preset physical constraint range, so as to ensure that the prediction results meet the actual operating conditions of flexible resources; truncation coverage loss assessment refers to the process of statistically analyzing the probability prediction interval after physical constraint truncation correction to evaluate the degree of reduction in prediction interval coverage caused by the truncation operation; adaptive interval width compensation refers to the dynamic adjustment of the prediction interval width after truncation correction based on the truncation coverage loss assessment results, so as to compensate for the loss of uncertainty information caused by boundary truncation during the physical constraint processing.

[0055] For example, the server first determines the corresponding physical constraints based on the type of flexible resource to be predicted, compares the prediction result to be constrained with the preset physical constraints, and retains the prediction result when it is within the preset physical constraints. When the prediction result is outside the preset physical constraints, the excess portion is truncated and corrected, adjusting the prediction result exceeding the preset physical constraints to the corresponding constraint range, thus obtaining the truncated and corrected intermediate prediction result. Since truncating and correcting may change the distribution characteristics of the original probability prediction result and may lead to inaccurate final prediction results, it is necessary to evaluate the truncated coverage loss rate of the intermediate prediction result after truncating and correcting. By calculating the actual coverage rate, the coverage loss caused by truncating is evaluated, the asymmetry of truncating is analyzed, and the truncated coverage loss rate evaluation result is obtained. Based on the truncated coverage loss rate evaluation result, the prediction interval width corresponding to the intermediate prediction result is adaptively adjusted, and the interval width is adaptively compensated in the direction that is not truncated, so that the truncated interval still maintains the target coverage rate, thus obtaining the baseline power probability prediction result of the flexible resource to be predicted. The final output probability interval maintains the target coverage rate (e.g., 85%) while satisfying the physical constraints.

[0056] In this embodiment, by determining the physical constraint type corresponding to the flexible resource type to be predicted, the prediction results of the constraints are subjected to truncation correction, truncation coverage loss rate assessment and adaptive interval width compensation, so that the prediction results maintain reasonable probability coverage capability while satisfying physical feasibility, thereby improving the accuracy, reliability and practical application value of the baseline power prediction results.

[0057] In an exemplary embodiment, the steps of truncating and correcting the prediction result based on preset physical constraints to obtain a truncated intermediate prediction result; evaluating the truncated coverage loss of the intermediate prediction result to obtain a truncated coverage loss evaluation result; and performing adaptive interval width compensation on the intermediate prediction result based on the truncated coverage loss evaluation result to obtain a baseline power probability prediction result for the flexible resource to be predicted can be performed by the following steps:

[0058] Based on the preset physical constraints for different resource types, corresponding truncation corrections are applied to the prediction results to be constrained:

[0059] 1) Physical constraints on photovoltaic resources:

[0060] Nighttime deterministic constraint: Based on local sunrise and sunset times, all predicted samples are corrected to zero during the nighttime period, while the interval width is compressed to zero. The coverage during the nighttime period is defined as 100%.

[0061] Sunrise and sunset Sigmoid-type smooth transition: A transition window is set before and after sunrise and sunset, using a Sigmoid-type transition function instead of hard truncation, so that the predicted samples smoothly transition from zero to normal values: for sunrise transition ( ): For sunset transition ( ): Where t is the current prediction time, and These are sunrise and sunset times, respectively. The transition window width is k, and the transition rate parameter is k (default). ), where s(t) is the transition coefficient. The predicted sample within the transition window is multiplied by the transition coefficient s(t) to smoothly transition the predicted value from zero to the normal value, avoiding unreasonable prediction results caused by abrupt changes in the interval.

[0062] Non-negativity constraint during the daytime: Correct prediction samples that are less than zero to zero.

[0063] Power cap constraint: The predicted sample exceeding the photovoltaic installed capacity is truncated to the installed capacity.

[0064] 2) Physical constraints on EV (Electric Vehicle) charging station resources:

[0065] Total capacity constraint for charging facilities: Predicted samples exceeding the total capacity of charging facilities within the charging station are truncated to the total capacity value.

[0066] Charging behavior time pattern constraint: Based on user behavior statistics, during periods with statistically no charging behavior, the prediction sample is narrowed towards zero. Specifically, the charging probability for each time period is calculated based on historical charging behavior statistics. During periods of low charging probability, the predicted samples are multiplied by a shrinkage factor. ,in As a retention factor, its theoretical value range is 0 < < 1, take it here = 0.1, This indicates a period of time with no charging activity in a statistical sense (extremely low probability of charging). The predicted value is taken as 10% of the original value, which avoids setting it to zero while significantly reducing the predicted value and the interval width. This balances the representation requirements of occasional charging behavior at EV charging stations with the shrinking constraint effect of low probability periods.

[0067] Non-negativity constraint: Corrects prediction samples that are less than zero to zero.

[0068] 3) Physical constraints on central air conditioning resources:

[0069] Temperature-power coupling constraint: There is a coupling relationship between the minimum power of the air conditioner and the outdoor temperature—when the outdoor temperature is below the start-up threshold, the air conditioner may not operate; when the outdoor temperature is above the start-up threshold, the air conditioner power increases with the temperature, and there is a minimum operating power. . Specifically:

[0070]

[0071] in This represents the minimum power rating for the air conditioner, where T is the outdoor temperature. To trigger the temperature threshold, To achieve the highest design temperature, This is the air conditioner's maximum operating power. It will be lower than... The predicted samples were corrected to (when T> (time), will be higher The predicted samples were truncated to .

[0072] Minimum operating time constraint: Once the air conditioner is started, it must maintain operation for at least [duration not specified]. During this period, power is not allowed to drop to a minimum. The following is achieved through curve volatility constraints: if continuous If the fluctuation range of the predicted sample exceeds the threshold within a given time period, the fluctuating portion will be smoothed to limit the rate of change of the curve to no more than the maximum permissible rate of change.

[0073]

[0074] Non-negativity constraint: Corrects prediction samples that are less than zero to zero.

[0075] Recalculate the actual coverage for the truncated N predicted samples and assess the coverage loss caused by the truncation:

[0076] Calculate the coverage of the original interval before truncation: ;

[0077] Calculate the coverage of the truncated interval: ;

[0078] Calculate the coverage loss: ;

[0079] Analyze the asymmetry of the cutoff: Calculate the lower bound cutoff loss separately. and upper bound cutoff loss .

[0080] Based on the coverage loss assessment results, the interval width is adaptively compensated in the untruncated direction to ensure that the truncated interval still maintains the target coverage:

[0081] ① Symmetrical truncation case ( ≈ ): Based on the truncated interval, the interval width is symmetrically expanded to both sides. The expansion is achieved by adjusting the quantile positions: the target quantile is adjusted from 7.5% / 92.5% to (7.5% - / 2)% / (92.5% + / 2)%, recalculate the quantiles on the original domain sample before truncation, and then truncate the new quantiles.

[0082] ② Asymmetric truncation case ( ≠ ): Expand the interval width only in the direction of greater coverage loss. For example, if the lower bound truncation loss is large (negative values ​​are truncated to zero, causing the lower bound to shift upward), then expand the interval width in the direction of the upper bound to compensate for the coverage loss caused by the lower bound truncation.

[0083] ③ Resource type-specific calibration strategy:

[0084] Photovoltaic resources: During the nighttime period, the interval width is compressed to zero (deterministic constraint), while during the daytime period, the interval width is dynamically adjusted based on the uncertainty of irradiance prediction. Within the sunrise-sunset transition window, the interval width smoothly transitions from zero to the normal width with the Sigmoid transition coefficient s(t).

[0085] EV charging station resources: The upper limit of the peak-hour range is subject to a hard capacity constraint (truncation to...). The lower bound is appropriately expanded based on the randomness of charging behavior to compensate for the coverage loss caused by the upper bound truncation. The width of the interval during low charging probability periods varies with the contraction factor. The corresponding contraction.

[0086] Central air conditioning resources: The range width is related to temperature forecast uncertainty; that is, during periods of high temperature forecast uncertainty, the air conditioning power forecast range width increases accordingly. During periods with minimum operating time constraints, the lower bound of the range is not lower than... The upper bound is expanded accordingly to compensate for the loss of coverage.

[0087] After the probability interval is recalibrated based on physical constraints, the final output probability interval maintains target coverage (e.g., 85%) while satisfying physical constraints, allowing schedulers to accurately assess and predict risks.

[0088] In this embodiment, physical constraint strategies are designed based on the specificity of resource type, rather than being uniformly processed. The physical constraint mechanism of zero output at night enhances the determinism of photovoltaic nighttime forecasts; the minimum operating time constraint of air conditioning enhances the temporal continuity of air conditioning forecasts, ensuring that the forecast results are within the actual physical operating range, thereby enhancing the accuracy and reliability of the baseline power forecast results.

[0089] In one exemplary embodiment, such as Figure 3 As shown, the pre-trained baseline energy probability prediction model can be obtained through steps 401 to 405:

[0090] Step 401: Obtain the training data for each sample's flexible resources. The training data includes historical operational sample data of the sample's flexible resources and related influencing factor sample data.

[0091] Among them, flexible resources for samples refer to flexible resource samples used to train the baseline electricity probability prediction model, which may include central air conditioning, electric vehicle charging stations, photovoltaic power generation, etc.

[0092] For example, the server obtains the corresponding historical operational sample data and associated influencing factor sample data based on the flexible resource type, wherein:

[0093] Photovoltaic resources: Historical operating data includes photovoltaic inverter power data; influencing factor sample data includes meteorological characteristics such as irradiance, temperature, humidity, and wind speed, as well as time characteristics such as month and hour.

[0094] Electric vehicle charging station resources: Historical operation data includes the total power of the charging station and the charging status data of each charging pile; sample data of influencing factors include meteorological characteristics such as temperature and rainfall, as well as time characteristics such as holidays, days of the week, and hours.

[0095] Central air conditioning resources: Historical operating data includes the start-stop status and power data of air conditioning equipment; sample data of influencing factors include meteorological characteristics such as outdoor temperature and humidity, as well as time characteristics such as month and hour. The air conditioning load has strong temperature-power coupling and continuous fluctuation.

[0096] Subsequently, the server constructs supervised learning samples according to the prediction task requirements, using historical data within a continuous time period as input features and the baseline electricity consumption for a relative future time period as the prediction target, thus forming the dataset required for model training. Samples of different resource types are distinguished by resource type labels to facilitate differentiated design of prior distributions and physical constraints in subsequent processes.

[0097] Step 402: Perform normal quantile transformation on each training data to obtain the corresponding target training data.

[0098] Among them, normal quantile transformation refers to a data distribution mapping method used to map the original data to an approximate standard normal distribution space.

[0099] For example, the server first identifies and removes outlier data from the training data, and then uses linear interpolation to fill in missing data. Subsequently, it standardizes each variable to make it meet the distribution form of zero mean and unit variance, thereby eliminating the influence of different units on model training. Then, the NQT method is used to perform distribution mapping on the preprocessed data. The cumulative probability corresponding to each data sample is obtained through the empirical cumulative distribution function. Then, the inverse cumulative distribution function of the standard normal distribution is used to map various data samples to the standard normal distribution space, so that the transformed data approximately follows the standard normal distribution, thereby obtaining the target training data.

[0100] Step 403: Determine the prior parameters of the model parameters of the baseline electricity probability prediction model to be trained according to the resource type of each sample flexible resource, and establish the prior distribution of the model parameters based on the prior parameters.

[0101] Here, model parameters refer to the trainable parameters corresponding to each network layer in the baseline electricity probability prediction model. In this embodiment, the model parameters are Bayesian random variables, and their values ​​follow the corresponding probability distribution. The posterior distribution of the model parameters is obtained through model training, and multiple sets of model parameters are extracted from the variational posterior distribution for probability prediction during the prediction stage. The prior distribution of the model parameters refers to the initial probability distribution corresponding to the model parameters before the start of model training, which is used to describe the initial uncertainty of the model parameters.

[0102] For example, the server determines the prior parameters of the model parameters for the baseline power probability prediction model to be trained based on the resource type of each sample flexible resource. Then, it establishes the prior distribution of the model parameters based on the prior parameters, so that the model has the corresponding probability distribution characteristics before training begins. Since flexible resources of different types have different operating rules, they each have corresponding preset prior parameters.

[0103] Step 404: Train the baseline electricity probability prediction model to be trained based on the target training data and the prior distribution of model parameters to construct the variational posterior distribution of model parameters.

[0104] Among them, the variational posterior distribution refers to the probability distribution of model parameters obtained by continuously updating the model parameters using the target training data, and is used to describe the uncertainty of model parameters after the model training is completed.

[0105] For example, the server inputs the target training data into the baseline power probability prediction model and trains the model by combining the prior distribution of the model parameters. This allows the model to gradually learn the mapping relationship between the historical operating data of the sample flexible resources and the associated influencing factors. As the model training continues, the variational posterior distribution corresponding to the model parameters is gradually constructed.

[0106] Step 405: Use variational inference to learn the variational posterior distribution of the model parameters and iteratively update the model parameters until the preset training termination condition is met, and obtain the trained baseline power probability prediction model.

[0107] Variational inference is an approximate inference method based on Bayesian theory, used to estimate the uncertainty of model parameters based on training data. This method constructs a variational posterior distribution that is close to the true posterior distribution and optimizes the parameters corresponding to the variational posterior distribution so that the variational posterior distribution can characterize the probability distribution features of the model parameters. Variational parameters refer to the parameters used to describe the variational posterior distribution of the model parameters.

[0108] For example, during model training, a variational inference method is used to train the model. By continuously adjusting the model parameters based on the target training data, the variational posterior distribution is optimized to gradually approximate the true posterior distribution, thus achieving the learning of the posterior distribution of the model parameters. During training, an objective function (EvidenceLowerBond, ELBO) is constructed, consisting of complexity loss and likelihood loss. Since NQT has transformed the data distribution to a standard normal domain aligned with the prior distribution, the initial value of the KL divergence (Kullback-Leibler Divergence, relative entropy) is small, making ELBO optimization easier to converge. To achieve differentiability of the random sampling process, a reparameterization technique is used to sample the parameters to obtain the final random samples. Expressed as mean Standard deviation With random noise Combination forms: ,in ~N(0,1) allows the gradient to propagate back through the sampling process. Since NQT approximates the standard normal distribution of data, reparameterized sampling is more stable in the standard normal domain, avoiding gradient explosion caused by excessive sampling variance in the skewed distribution domain. In this embodiment, the Adam (Adaptive Moment Estimation) optimization algorithm is used to iteratively update the variational parameters corresponding to the variational posterior distribution. The learning rate can be set to 0.001, the batch size to 32, and the number of training rounds to 200. After multiple rounds of iterative training, the optimal posterior distribution of the model parameters is obtained, and the posterior mean and posterior variance corresponding to the weights and biases of each layer are saved, resulting in the trained baseline electricity probability prediction model. The NQT transformation parameters are fixed in the early stages of training and can be fine-tuned in the later stages based on the convergence of the variational posterior to further optimize the alignment between the data distribution and the prior distribution.

[0109] In the aforementioned method for predicting the baseline electricity consumption of flexible resources, the model training and prediction phases employ identical outlier handling, missing value completion, standardization, and normal quantile transformation to ensure consistent data distribution characteristics between the training and prediction data. By performing normal quantile transformation on the training data, the data distribution is aligned with the prior distribution, reducing the impact of skewed distributions and extreme values ​​on the training process, improving model convergence speed and generalization ability, while ensuring the statistical validity of the probability interval within the original domain. A prior distribution of model parameters is established based on the flexible resource types in the samples, and training is performed using the target training data, enabling the model to integrate the operational characteristics of different types of flexible resources and improving its adaptability to different resource types. By employing variational inference to learn the variational posterior distribution of model parameters, the model parameters are expanded from deterministic parameters to a probabilistic distribution, quantifying the uncertainty of the model parameters and providing a model foundation for the prediction phase. By incorporating the transformation parameters and variational posterior parameters of NQT into a unified optimization framework, the NQT and prior distribution co-optimization mechanism aligns the data distribution with the prior distribution, reduces the initial value of KL divergence, accelerates ELBO convergence, and improves the stability of reparameterized sampling.

[0110] In an exemplary embodiment, the peak-plate prior parameters of the model parameters of the baseline power probability prediction model to be trained are determined according to the resource type of each sample flexible resource, and the prior distribution of the model parameters is established based on the peak-plate prior parameters.

[0111] The Spike-and-Slab prior is a probabilistic prior distribution used for sparsifying Bayesian model parameters. It consists of two parts: a spike distribution and a slab distribution. The spike part represents the probabilistic features where the model parameters are close to zero, while the slab part represents the probabilistic features where the model parameters have larger values ​​and retain effective feature information.

[0112] For example, the server determines the corresponding prior distribution parameters based on different types of flexible resources, where the Spike-and-Slab prior consists of a mixture of two Gaussian distributions:

[0113]

[0114] in For mixed weights, The standard deviation of the peaked distribution. Let be the standard deviation of the flat plate distribution. The prior parameters for different resource types are designed as follows:

[0115] For photovoltaic resources, an enhanced sparsity shift prior is adopted: =0.5, =10, =0.8. Photovoltaic output exhibits strong sparsity (zero at night, non-zero during the day). The smaller peak standard deviation (0.5 instead of 1) enhances the prior weight in the zero-value region, causing the dense layer parameters to tend to cluster near zero, consistent with the physical constraint mechanism of zero photovoltaic output at night. Higher mixing weight. Setting the weight to 0.8 constrains more parameters to a peaked distribution, enhancing the sparsity-inducing effect. Mixed weights Set hyperparameters prior ~Beta(8,2) is automatically learned to be optimal during training through variational inference. value.

[0116] For electric vehicle resources, a time-modulated prior is adopted: =1, =10, Dynamically adjusted according to time characteristics. EV charging behavior exhibits temporal clustering (weekday morning and evening peak hours), with a mixed weighting of flat plate distribution (1- Dynamically adjusted based on time characteristics (weekdays, hours): Peak hours (1- Increase (allows for more weighted parameters to adapt to rapid changes in charging power), during off-peak hours (1- This reduces (enhances sparsity, adapting to low-power periods). Specifically, ,in For time modulation functions based on historical charging behavior statistics, =0.5, ∈ Mixed weights Set hyperparameters prior ~Beta(5,5) is automatically learned as the optimal value during training through variational inference. value.

[0117] For central air conditioning resources, a continuous enhancement prior is adopted: =1, =5, =0.3. Air conditioning load is strongly coupled with outdoor temperature and exhibits continuous fluctuations. A smaller flat standard deviation (5 instead of 10) avoids excessive sparsity, retaining more moderately weighted parameters to adapt to the continuous variation characteristics of air conditioning load. Lower mixed weights. =0.3 constrains more parameters to the flat plate distribution, allowing for a wider range of parameter variations. Mixed weights Set hyperparameters prior ~Beta(3,7) is automatically learned to be optimal during training through variational inference. value.

[0118] After determining the corresponding Spike-and-Slab prior parameters for flexible resources of different resource types, the prior distributions corresponding to the model parameters are established.

[0119] In this embodiment, the resource type adaptive prior design exhibits a synergistic effect: the offset Spike-and-Slab prior for photovoltaics and the physical constraint of zero photovoltaic output at night are mechanistically consistent, i.e., the prior distribution tends towards zero, and the physical constraint also forces zero output at night. Together, they enhance the determinism of photovoltaic nighttime predictions. Similarly, the continuity enhancement prior for air conditioning and the minimum operating time constraint for air conditioning are synergistic, i.e., the prior retains continuity parameters, and the physical constraint limits curve volatility. Together, they enhance the temporal continuity of air conditioning predictions, allowing for a more accurate description of the uncertainty of model parameters and improving the reliability of baseline power prediction results. By adopting a resource type adaptive hybrid prior, prior parameters are designed differently based on resource operating characteristics, improving the parameter sampling efficiency and model training stability for each resource type, and mitigating the problems of excessive contraction or convergence difficulties that may result from a single Gaussian prior.

[0120] In one exemplary embodiment, such as Figure 4 As shown, the target training data is input into the baseline power probability prediction model to be trained. The target training data is trained sequentially based on the input layer, one-dimensional convolutional layer, max pooling layer, bidirectional long short-term memory network layer, and dense output layer of the baseline power probability prediction model, combined with the prior distribution of the model parameters. Specifically, the input layer receives the target training data and inputs it into the one-dimensional convolutional layer; the one-dimensional convolutional layer is used to extract local features from the target training data; the max pooling layer is used to compress the extracted features; the bidirectional long short-term memory network layer is used to learn temporal features from the compressed features; and the dense output layer is used to output the predicted Gaussian distribution parameters.

[0121] The input layer, one-dimensional convolutional layer, and max pooling layer together constitute a convolutional neural network (CNN); the bidirectional long short-term memory (BiLSTM) network layer is a recurrent neural network structure that can simultaneously learn historical temporal information and future temporal correlation information, and is used to learn temporal features from pooled features; the ReLU function (Rectified Linear Unit) is a non-linear activation function used to perform non-linear mapping on the feature results after convolution operation, so as to enhance the model's ability to express complex non-linear relationships.

[0122] For example, the baseline energy prediction model employs a network structure combining a convolutional neural network and a bidirectional long short-term memory network, consisting of an input layer, a one-dimensional convolutional layer, a max-pooling layer, a bidirectional long short-term memory network layer, and a dense output layer. After the target training data is input into the model, the one-dimensional convolutional layer first performs convolution operations to extract local features from the target training data. The ReLU function is then used as the activation function to perform a non-linear mapping on the convolution operation results, thereby obtaining local feature information that characterizes the variation pattern of the baseline energy of flexible resources. Next, the features output from the one-dimensional convolutional layer are input into the max-pooling layer. The max-pooling layer performs feature compression processing on the local features, reducing the feature dimensionality while retaining key features representative of the baseline energy variation, thus improving the model's robustness to noise disturbances and local changes. Subsequently, the compressed features processed by the max-pooling layer are input into a bidirectional long short-term memory (LSS) network layer. In this network, the forward LSS learns feature changes along the historical time axis, while the backward LSS extracts correlation information along the reverse time axis, thus fully exploring the long-term temporal features and complex dynamic patterns in the operation of flexible resources. Finally, the temporal features output from the BLS layer are input into a dense output layer. This layer maps the input feature data to a predetermined number of output neurons and generates parameters describing the baseline energy probability distribution based on the output results. This obtains the uncertainty information corresponding to the baseline energy prediction results and generates the final output. The weight parameters in the one-dimensional convolutional layer and the BLS layer use a standard Gaussian distribution as the prior distribution. .

[0123] In this embodiment, the one-dimensional convolutional layer and the bidirectional long short-term memory network layer are responsible for feature extraction and temporal modeling. Their parameter priors remain consistent, ensuring that the model's basic feature learning ability is not affected by differences in resource types. NQT transforms the target training data to the standard normal distribution domain, aligning it with the standard Gaussian prior distribution of the one-dimensional convolutional layer / bidirectional long short-term memory network layer and the Spike-and-Slab prior distribution of the dense output layer. This results in a smaller initial value of the KL divergence between the variational posterior and the prior, making ELBO optimization easier to converge and improving the stability of Bayesian parameter learning.

[0124] In summary, the flexible resource-based energy probability prediction method provided in this application can achieve at least the following technical effects:

[0125] 1) Provides a reliable probability interval for the baseline electricity, outputs probability prediction results at a certain confidence level, and ensures that the interval still maintains the target coverage after truncation through a probability interval recalibration mechanism that is aware of physical constraints, thus solving the problem that traditional truncation correction operations destroy the statistical validity of the interval.

[0126] 2) Simultaneously modeling two types of uncertainty, quantifying cognitive uncertainty through Bayesian inference (variational inference) and quantifying stochastic uncertainty through output distribution, resulting in more comprehensive and reliable predictions. The NQT-prior distribution co-optimization mechanism aligns the data distribution with the prior distribution, reduces the initial value of KL divergence, accelerates ELBO convergence, and improves the stability of reparameterized sampling.

[0127] 3) Differentiated processing for various types of heterogeneous and flexible resources: Prior distribution parameters and physical constraint strategies are designed based on the specificity of resource types, rather than uniform processing. The photovoltaic offset prior (the photovoltaic offset Spike-and-Slab prior) is consistent with the physical constraint mechanism of zero output at night, and the two work together to enhance the determinism of photovoltaic nighttime prediction; the air conditioning continuity enhancement prior works together with the minimum operating time constraint to enhance the temporal continuity of air conditioning prediction.

[0128] 4) Improved sampling efficiency: By adopting a hybrid prior with resource type adaptation, prior parameters are designed differently according to the resource operation characteristics, which improves the parameter sampling efficiency and model training stability of each resource type and alleviates the problems of excessive shrinkage or convergence difficulties that may be caused by traditional single Gaussian prior.

[0129] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0130] Based on the same inventive concept, this application also provides a flexible resource reference power probability prediction device for implementing the flexible resource reference power probability prediction method described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the flexible resource reference power probability prediction device provided below can be found in the limitations of the flexible resource reference power probability prediction method described above, and will not be repeated here.

[0131] In one exemplary embodiment, such as Figure 5As shown, a flexible resource-based energy probability prediction device is provided, comprising: a data acquisition module 10, a quantile transformation module 20, a probability prediction module 30, and a result generation module 40, wherein:

[0132] The data acquisition module 10 is used to acquire the forecast data of the flexible resource to be predicted based on the resource type of the flexible resource to be predicted; the forecast data includes the historical operation data of the flexible resource to be predicted and the related influencing factor data.

[0133] Quantile transformation module 20 is used to process the data to be predicted based on normal quantile transformation to obtain the target prediction data.

[0134] The probability prediction module 30 is used to extract multiple sets of model parameters from the posterior distribution of the model parameters of the pre-trained benchmark power probability prediction model, and perform a complete forward propagation on each set of model parameters in combination with the target prediction data to obtain the prediction Gaussian distribution corresponding to each set of model parameters.

[0135] The result generation module 40 is used to generate corresponding initial prediction results based on each predicted Gaussian distribution, and to process each initial prediction result based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline power probability prediction result of the flexible resource to be predicted.

[0136] In an exemplary embodiment, the result generation module 40 is further configured to process each prediction initial result based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the constrained prediction result of the baseline energy of the flexible resource to be predicted; and to perform physical constraint recalibration on the constrained prediction result according to the preset physical constraints corresponding to the flexible resource to be predicted to obtain the baseline energy probability prediction result of the flexible resource to be predicted.

[0137] In an exemplary embodiment, the result generation module 40 is further configured to determine the corresponding preset physical constraints based on the resource type of the flexible resource to be predicted; truncate and correct the prediction result to be constrained based on the preset physical constraints to obtain the intermediate prediction result after truncation correction; evaluate the truncation coverage loss of the intermediate prediction result to obtain the truncation coverage loss evaluation result; and perform adaptive interval width compensation on the intermediate prediction result based on the truncation coverage loss evaluation result to obtain the baseline power probability prediction result of the flexible resource to be predicted.

[0138] In an exemplary embodiment, the probability prediction module 30 is further configured to acquire training data for each sample flexible resource, including historical operational sample data of the sample flexible resource and associated influencing factor sample data; perform normal quantile transformation on each training data to obtain corresponding target training data; determine the prior parameters of the model parameters of the benchmark electricity probability prediction model to be trained according to the resource type of each sample flexible resource, and establish the prior distribution of the model parameters based on the prior parameters; train the benchmark electricity probability prediction model to be trained based on the target training data and the prior distribution of the model parameters to construct the variational posterior distribution of the model parameters; use variational inference to learn the variational posterior distribution of the model parameters, and iteratively update the model parameters until the preset training termination condition is met, thereby obtaining the trained benchmark electricity probability prediction model.

[0139] In an exemplary embodiment, the probability prediction module 30 is further configured to determine the peak-plate prior parameters of the model parameters of the benchmark power probability prediction model to be trained based on the resource type of each sample flexible resource, and to establish the prior distribution of the model parameters based on the peak-plate prior parameters.

[0140] In an exemplary embodiment, the probability prediction module 30 is further configured to input the target training data into the benchmark power probability prediction model to be trained, so as to train the target training data sequentially based on the input layer, one-dimensional convolutional layer, max pooling layer, bidirectional long short-term memory network layer, and dense output layer of the benchmark power probability prediction model to be trained, combined with the prior distribution of the model parameters; wherein: the input layer is used to receive the target training data and input the target training data into the one-dimensional convolutional layer; the one-dimensional convolutional layer is used to extract local features from the target training data; the max pooling layer is used to perform feature compression processing on the extraction results obtained by the local feature extraction; the bidirectional long short-term memory network layer is used to perform temporal feature learning on the compressed features after feature compression processing; and the dense output layer is used to output the predicted Gaussian distribution parameters.

[0141] The modules in the aforementioned flexible resource-based electricity probability prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0142] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores historical operational data of flexible resources and related influencing factors. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a flexible resource baseline energy probability prediction method.

[0143] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0144] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the following steps: obtaining prediction data for the flexible resource to be predicted based on the resource type of the flexible resource to be predicted; the prediction data includes historical operating data of the flexible resource to be predicted and associated influencing factor data; processing the prediction data based on normal quantile transformation to obtain target prediction data; extracting multiple sets of model parameters from the posterior distribution of model parameters of a pre-trained baseline energy probability prediction model, and performing a complete forward propagation on each set of model parameters in conjunction with the target prediction data to obtain the prediction Gaussian distribution corresponding to each set of model parameters; generating corresponding initial prediction results based on each prediction Gaussian distribution, and processing each initial prediction result based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline energy probability prediction result of the flexible resource to be predicted.

[0145] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0146] Each initial prediction result is processed based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the constrained prediction result of the baseline energy of the flexible resource to be predicted; according to the preset physical constraints corresponding to the flexible resource to be predicted, the physical constraint recalibration is performed on the constrained prediction result to obtain the baseline energy probability prediction result of the flexible resource to be predicted.

[0147] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0148] Determine the corresponding preset physical constraints based on the resource type of the flexible resource to be predicted; truncate and correct the prediction results based on the preset physical constraints to obtain the intermediate prediction results after truncation and correction; evaluate the truncation coverage loss of the intermediate prediction results to obtain the truncation coverage loss evaluation results.

[0149] Based on the assessment results of the coverage loss due to truncation, the intermediate prediction results are adaptively compensated for the interval width to obtain the baseline power probability prediction results of the flexible resources to be predicted.

[0150] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0151] The process involves acquiring training data for each sample of flexible resources, including historical operational data of the flexible resources and related influencing factor data; performing normal quantile transformation on each training data set to obtain the corresponding target training data; determining the prior parameters of the model parameters for the baseline electricity probability prediction model based on the resource type of each sample of flexible resources, and establishing a prior distribution of the model parameters based on the prior parameters; training the baseline electricity probability prediction model based on the target training data and the prior distribution of the model parameters to construct the variational posterior distribution of the model parameters; using variational inference to learn the variational posterior distribution of the model parameters and iteratively updating the model parameters until the preset training termination condition is met, thus obtaining the trained baseline electricity probability prediction model.

[0152] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0153] Based on the resource type of each sample's flexible resources, the peak-plate prior parameters of the model parameters of the baseline electricity probability prediction model to be trained are determined, and the prior distribution of the model parameters is established based on the peak-plate prior parameters.

[0154] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0155] The target training data is input into the baseline power probability prediction model to be trained. The model is then trained sequentially based on the input layer, one-dimensional convolutional layer, max pooling layer, bidirectional long short-term memory network layer, and dense output layer of the baseline power probability prediction model, combined with the prior distribution of the model parameters. Specifically: the input layer receives the target training data and inputs it into the one-dimensional convolutional layer; the one-dimensional convolutional layer is used to extract local features from the target training data; the max pooling layer is used to compress the extracted features; the bidirectional long short-term memory network layer is used to learn temporal features from the compressed features; and the dense output layer outputs the predicted Gaussian distribution parameters.

[0156] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0157] Based on the resource type of the flexible resource to be predicted, the prediction data of the flexible resource to be predicted is obtained. The prediction data includes the historical operation data of the flexible resource to be predicted and the related influencing factor data. The prediction data is processed based on the normal quantile transformation to obtain the target prediction data. Multiple sets of model parameters are extracted from the posterior distribution of the model parameters of the pre-trained baseline power probability prediction model, and a complete forward propagation is performed on each set of model parameters in combination with the target prediction data to obtain the prediction Gaussian distribution corresponding to each set of model parameters. The corresponding initial prediction results are generated according to each prediction Gaussian distribution, and each initial prediction result is processed based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline power probability prediction result of the flexible resource to be predicted.

[0158] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0159] Each initial prediction result is processed based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the constrained prediction result of the baseline energy of the flexible resource to be predicted; according to the preset physical constraints corresponding to the flexible resource to be predicted, the physical constraint recalibration is performed on the constrained prediction result to obtain the baseline energy probability prediction result of the flexible resource to be predicted.

[0160] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0161] The corresponding preset physical constraints are determined based on the resource type of the flexible resource to be predicted; the prediction results of the constraints are truncated and corrected based on the preset physical constraints to obtain the intermediate prediction results after truncation correction; the truncation coverage loss is evaluated on the intermediate prediction results to obtain the truncation coverage loss evaluation results; based on the truncation coverage loss evaluation results, the intermediate prediction results are adaptively compensated for the interval width to obtain the baseline power probability prediction results of the flexible resource to be predicted.

[0162] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0163] The process involves acquiring training data for each sample of flexible resources, including historical operational data of the flexible resources and related influencing factor data; performing normal quantile transformation on each training data set to obtain the corresponding target training data; determining the prior parameters of the model parameters for the baseline electricity probability prediction model based on the resource type of each sample of flexible resources, and establishing a prior distribution of the model parameters based on the prior parameters; training the baseline electricity probability prediction model based on the target training data and the prior distribution of the model parameters to construct the variational posterior distribution of the model parameters; using variational inference to learn the variational posterior distribution of the model parameters and iteratively updating the model parameters until the preset training termination condition is met, thus obtaining the trained baseline electricity probability prediction model.

[0164] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0165] Based on the resource type of each sample's flexible resources, the peak-plate prior parameters of the model parameters of the baseline electricity probability prediction model to be trained are determined, and the prior distribution of the model parameters is established based on the peak-plate prior parameters.

[0166] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0167] The target training data is input into the baseline power probability prediction model to be trained. The model is then trained sequentially based on the input layer, one-dimensional convolutional layer, max pooling layer, bidirectional long short-term memory network layer, and dense output layer of the baseline power probability prediction model, combined with the prior distribution of the model parameters. Specifically: the input layer receives the target training data and inputs it into the one-dimensional convolutional layer; the one-dimensional convolutional layer is used to extract local features from the target training data; the max pooling layer is used to compress the extracted features; the bidirectional long short-term memory network layer is used to learn temporal features from the compressed features; and the dense output layer outputs the predicted Gaussian distribution parameters.

[0168] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0169] Based on the resource type of the flexible resource to be predicted, the prediction data of the flexible resource to be predicted is obtained. The prediction data includes the historical operation data of the flexible resource to be predicted and the related influencing factor data. The prediction data is processed based on the normal quantile transformation to obtain the target prediction data. Multiple sets of model parameters are extracted from the posterior distribution of the model parameters of the pre-trained baseline power probability prediction model, and a complete forward propagation is performed on each set of model parameters in combination with the target prediction data to obtain the prediction Gaussian distribution corresponding to each set of model parameters. The corresponding initial prediction results are generated according to each prediction Gaussian distribution, and each initial prediction result is processed based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline power probability prediction result of the flexible resource to be predicted.

[0170] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0171] Each initial prediction result is processed based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the constrained prediction result of the baseline energy of the flexible resource to be predicted; according to the preset physical constraints corresponding to the flexible resource to be predicted, the physical constraint recalibration is performed on the constrained prediction result to obtain the baseline energy probability prediction result of the flexible resource to be predicted.

[0172] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0173] The corresponding preset physical constraints are determined based on the resource type of the flexible resource to be predicted; the prediction results of the constraints are truncated and corrected based on the preset physical constraints to obtain the intermediate prediction results after truncation correction; the truncation coverage loss is evaluated on the intermediate prediction results to obtain the truncation coverage loss evaluation results; based on the truncation coverage loss evaluation results, the intermediate prediction results are adaptively compensated for the interval width to obtain the baseline power probability prediction results of the flexible resource to be predicted.

[0174] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0175] The process involves acquiring training data for each sample of flexible resources, including historical operational data of the flexible resources and related influencing factor data; performing normal quantile transformation on each training data set to obtain the corresponding target training data; determining the prior parameters of the model parameters for the baseline electricity probability prediction model based on the resource type of each sample of flexible resources, and establishing a prior distribution of the model parameters based on the prior parameters; training the baseline electricity probability prediction model based on the target training data and the prior distribution of the model parameters to construct the variational posterior distribution of the model parameters; using variational inference to learn the variational posterior distribution of the model parameters and iteratively updating the model parameters until the preset training termination condition is met, thus obtaining the trained baseline electricity probability prediction model.

[0176] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0177] Based on the resource type of each sample's flexible resources, the peak-plate prior parameters of the model parameters of the baseline electricity probability prediction model to be trained are determined, and the prior distribution of the model parameters is established based on the peak-plate prior parameters.

[0178] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0179] The target training data is input into the baseline power probability prediction model to be trained. The model is then trained sequentially based on the input layer, one-dimensional convolutional layer, max pooling layer, bidirectional long short-term memory network layer, and dense output layer of the baseline power probability prediction model, combined with the prior distribution of the model parameters. Specifically: the input layer receives the target training data and inputs it into the one-dimensional convolutional layer; the one-dimensional convolutional layer is used to extract local features from the target training data; the max pooling layer is used to compress the extracted features; the bidirectional long short-term memory network layer is used to learn temporal features from the compressed features; and the dense output layer outputs the predicted Gaussian distribution parameters.

[0180] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0181] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0182] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0183] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A flexible resource-based baseline energy probability prediction method, characterized in that, The method includes: Based on the resource type of the flexible resource to be predicted, obtain the data to be predicted for the flexible resource to be predicted; the data to be predicted includes the historical operating data of the flexible resource to be predicted and the associated influencing factor data. The data to be predicted is processed based on the normal quantile transformation to obtain the target prediction data; Multiple sets of model parameters are extracted from the posterior distribution of model parameters of the pre-trained baseline power probability prediction model, and a complete forward propagation is performed on each set of model parameters in combination with the target prediction data to obtain the prediction Gaussian distribution corresponding to each set of model parameters. Based on the predicted Gaussian distribution, corresponding initial prediction results are generated. Each initial prediction result is then processed based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline power probability prediction result of the flexible resource to be predicted.

2. The method according to claim 1, characterized in that, The process of processing each of the initial prediction results based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline energy probability prediction result of the flexible resource to be predicted includes: The initial prediction results are processed based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the constrained prediction results of the baseline electricity of the flexible resource to be predicted. Based on the preset physical constraints corresponding to the flexible resource to be predicted, the prediction result to be constrained is recalibrated to obtain the baseline power probability prediction result of the flexible resource to be predicted.

3. The method according to claim 2, characterized in that, The step of recalibrating the prediction result to be constrained based on the preset physical constraints corresponding to the flexible resource to be predicted, to obtain the baseline power probability prediction result of the flexible resource to be predicted, includes: Determine the corresponding preset physical constraints based on the resource type of the flexible resource to be predicted; Based on the preset physical constraints, the prediction result to be constrained is truncated and corrected to obtain the intermediate prediction result after truncation and correction. The intermediate prediction results are subjected to a truncation coverage loss assessment to obtain the truncation coverage loss assessment result; Based on the assessment results of the cutoff coverage loss, adaptive interval width compensation is applied to the intermediate prediction results to obtain the baseline power probability prediction results of the flexible resources to be predicted.

4. The method according to claim 1, characterized in that, The pre-trained baseline energy probability prediction model was obtained through training in the following manner: Acquire the training data for each sample's flexible resources, including historical operational sample data and associated influencing factor sample data of the sample's flexible resources; Perform normal quantile transformation on each of the training data to obtain the corresponding target training data; Based on the resource type of each sample flexible resource, the prior parameters of the model parameters of the benchmark electricity probability prediction model to be trained are determined, and the prior distribution of the model parameters is established based on the prior parameters. The baseline power probability prediction model to be trained is trained based on the target training data and the prior distribution of the model parameters to construct the variational posterior distribution of the model parameters. Variational inference is used to learn the variational posterior distribution of the model parameters, and the model parameters are iteratively updated until the preset training termination condition is met, thus obtaining the trained baseline power probability prediction model.

5. The method according to claim 4, characterized in that, The step of determining the prior parameters of the model parameters of the baseline electricity probability prediction model to be trained based on the resource type of each of the sample flexible resources, and establishing the prior distribution of the model parameters based on the prior parameters, includes: Based on the resource type of each sample flexible resource, the peak-plate prior parameters of the model parameters of the benchmark power probability prediction model to be trained are determined, and the prior distribution of the model parameters is established based on the peak-plate prior parameters.

6. The method according to claim 4, characterized in that, The training of the baseline electricity probability prediction model based on the target training data and the prior distribution of the model parameters includes: The target training data is input into the baseline battery probability prediction model to be trained, and the model is trained sequentially based on the input layer, one-dimensional convolutional layer, max pooling layer, bidirectional long short-term memory network layer, and dense output layer of the baseline battery probability prediction model, combined with the prior distribution of the model parameters; wherein: The input layer is used to receive the target training data and input the target training data into the one-dimensional convolutional layer; The one-dimensional convolutional layer is used to extract local features from the target training data; The max pooling layer is used to perform feature compression processing on the extraction results obtained from local feature extraction. The bidirectional long short-term memory network layer is used to perform temporal feature learning on the compressed features after feature compression processing. The dense output layer is used to output the predicted Gaussian distribution parameters.

7. A flexible resource-based energy probability prediction device, characterized in that, The device includes: The data acquisition module is used to acquire the forecast data of the flexible resource to be predicted based on the resource type of the flexible resource to be predicted; the forecast data includes the historical operation data of the flexible resource to be predicted and the associated influencing factor data. The quantile transformation module is used to process the data to be predicted based on the normal quantile transformation to obtain the target prediction data; The probability prediction module is used to extract multiple sets of model parameters from the posterior distribution of model parameters of the pre-trained benchmark power probability prediction model, and perform a complete forward propagation on each set of model parameters in combination with the target prediction data to obtain the prediction Gaussian distribution corresponding to each set of model parameters. The result generation module is used to generate corresponding initial prediction results based on each of the predicted Gaussian distributions, and to process each of the initial prediction results based on the inverse normal quantile transformation corresponding to the normal quantile transformation to obtain the baseline power probability prediction result of the flexible resource to be predicted.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images