A regional-level new energy peak shaving demand analysis method and system

Abstract

This invention discloses a regional-level renewable energy peak-shaving demand analysis method and system. Based on an input feature vector reflecting the power system's operating conditions, a trained base model pool is used to obtain correction coefficients for the minimum technical output of conventional power sources and tie-line switching power. These correction coefficients are then used to adjust planned values. A block resampling strategy is used to construct multiple training datasets, and each training subset is used to train an extreme learning machine model. The total prediction error of the multiple trained extreme learning machine models is obtained based on statistical cumulative factors. n The cumulative output is obtained by using Gram-Charlier A series expansion to obtain the probability density function of electricity load and total regional renewable energy output. Based on the modified minimum technical output of conventional power sources, the modified tie-line switching power, and the two probability density functions, the probability distribution of peak-shaving demand is calculated using a Gaussian mixture model, thereby improving the accuracy and efficiency of peak-shaving demand analysis.

Description

Technical Field

[0001] This invention relates to the field of renewable energy peak-shaving demand analysis technology, and in particular to a regional-level renewable energy peak-shaving demand analysis method and system. Background Technology

[0002] In recent years, the penetration rate of new energy sources, represented by offshore wind power and distributed photovoltaic power, in the power system has been continuously increasing. The inherent intermittency (such as the anti-peak-shaving characteristics of wind power) and volatility of their output have significantly increased the uncertainty of the power grid. In the planning and operation of new power systems, accurately calculating the peak-shaving resource requirements needed to balance such fluctuations and providing comprehensive probability distribution information and safety boundaries are key links to ensure the safety and stability of the system and promote the efficient consumption of new energy sources.

[0003] Traditional power grid regional renewable energy (including various renewable energy types within a designated area, such as offshore wind power, onshore wind power, distributed photovoltaic, centralized photovoltaic, etc.) peak-shaving demand analysis can determine the installed capacity margin and absorption capacity of renewable energy within a designated area based on calculation results. It is generally based on the following formula: Regional renewable energy peak-shaving demand = Electricity load Minimum technical output of conventional power supplies Tether line switching power (input positive) New energy power output. Traditional methods for calculating demand are generally based on planned value estimation and point value calculation. The variables in the formula are treated as fixed point values, and numerical scaling and arithmetic addition and subtraction are performed according to the proportion of system installation on different dates to obtain a deterministic peak-shaving power gap.

[0004] However, with the rapid increase in system uncertainty, the limitations of this traditional calculation method have become increasingly apparent. First, the method lacks risk quantification capabilities, outputting only a single numerical value that cannot characterize the potential fluctuation range and probability distribution of peak-shaving demand, making it difficult to support modern power systems' dispatch decisions based on risk warning and probabilistic optimization. Second, the input benchmark of the calculation method contains systematic errors. Key parameters in the formula, such as the minimum technical output of conventional power sources and the planned output of tie lines, are usually determined by simple proportional estimation based on historical and future planned values, failing to fully consider the impact of complex operating scenarios such as real-time unit combination and network congestion, resulting in inherent biases at the source of the calculation. Finally, the calculation method is inefficient. Existing research and practice, with the few methods attempting to consider uncertainty, generally rely on traditional methods such as Monte Carlo simulation when merging multi-source probability distributions such as load and renewable energy output. The enormous computational and time complexity of these methods makes it difficult to meet the timeliness requirements of online rolling optimization and real-time dispatch. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a regional-level new energy peak-shaving demand analysis method and system, which can improve the accuracy and efficiency of peak-shaving demand analysis.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A regional-level renewable energy peak-shaving demand analysis method includes the following steps: Based on the obtained input feature vectors reflecting the operating conditions of the power system, the correction coefficients for the minimum technical output of conventional power sources and the switching power of tie lines are obtained using the trained base model pool. Obtain planned values ​​for the minimum technical output of conventional power supplies and the switching power of tie lines; The planned value is corrected using the correction factor to obtain the corrected minimum technical output of conventional power and the corrected tie-line switching power. Multiple training datasets were constructed using a block resampling strategy based on historical electricity load time-series data and historical regional total renewable energy output time-series data. Train an extreme learning machine model using each training subset in the training dataset to obtain multiple trained extreme learning machine models. The total prediction error of the multiple trained Extreme Learning Machine models is obtained based on the cumulative effect in statistics. n The order of cumulative quantity, and based on the total prediction error. n The cumulative quantities are expanded using Gram-Charlier A series to obtain the first probability density function of the electricity load and the second probability density function of the total output of regional new energy sources. The peak demand probability distribution is calculated using a Gaussian mixture model based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the first probability density function, and the second probability density function.

[0007] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A regional-level renewable energy peak-shaving demand analysis system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various steps in the aforementioned regional-level renewable energy peak-shaving demand analysis method.

[0008] The beneficial effects of this invention are as follows: Based on the obtained input feature vectors reflecting the power system's operating conditions, correction coefficients for the minimum technical output of conventional power sources and tie-line exchange power are obtained using the trained base model pool. These correction coefficients are then used to adjust the planned values ​​of the minimum technical output of conventional power sources and tie-line exchange power. An ensemble learning strategy is introduced to effectively correct systematic deviations in the planned values ​​of the minimum technical output of conventional power sources and tie-line exchange power. Next, multiple training datasets are constructed using a block resampling strategy based on historical electricity load time-series data and historical regional total renewable energy output time-series data. Each training subset is used to train an extreme learning machine model. The total prediction error of the multiple trained extreme learning machine models is obtained based on statistical cumulative factors. n The order cumulative, and based on the total prediction error n The cumulative quantities are expanded using Gram-Charlier A series to obtain the first probability density function of the electricity load and the second probability density function of the total output of regional renewable energy. This preserves the time-series fluctuation characteristics of renewable energy output. By synthesizing cumulative quantities, data and model uncertainties are captured, overcoming the limitations of traditional methods that are based on assumptions about a specific probability distribution. Finally, based on the modified minimum technical output of conventional power sources, the modified tie-line switching power, the first probability density function, and the second probability density function, a Gaussian mixture model is used to calculate the probability distribution of peak-shaving demand. The use of a Gaussian mixture model to replace the traditional Monte Carlo simulation transforms the complex probability fusion process into efficient matrix algebra operations, significantly reducing computational and time complexity, thereby improving the accuracy and efficiency of peak-shaving demand analysis. Attached Figure Description

[0009] Figure 1 This is a flowchart of a regional-level renewable energy peak-shaving demand analysis method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a regional-level new energy peak-shaving demand analysis system according to an embodiment of the present invention. Detailed Implementation

[0010] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0011] Before detailing the embodiments of this application, some related concepts will first be explained: Random Forest (RF): A classifier that uses multiple trees to train and predict samples; Lightweight Gradient Boosting Machine (LightGBM): A lightweight gradient boosting machine learning framework designed to efficiently process large-scale data. It is an improved version of the GBDT (Gradient Boosting Decision Tree) algorithm, focusing on improving training speed, reducing memory consumption, and maintaining high accuracy. Multilayer Perceptron (MLP): A feedforward artificial neural network model that learns complex nonlinear relationships by mapping input data to output data; Extreme Learning Machine (ELM): A type of machine learning system or method based on Feedforward Neuron Network (FNN), suitable for both supervised and unsupervised learning problems; Gram-Charlier A series: A mathematical tool for probability distribution analysis, mainly used to expand the probability density function of a random variable into a series form, where each term is expressed based on the cumulants (also known as semi-invariants) of the random variable; Gaussian Mixture Model (GMM): It uses the Gaussian probability density function (normal distribution curve) to accurately quantify things. It is a model that decomposes things into several components based on the Gaussian probability density function (normal distribution curve).

[0012] In existing technologies, traditional regional-level renewable energy peak-shaving demand analysis methods for power grids lack risk quantification capabilities. These methods only output a single numerical value, failing to characterize the potential fluctuation range and probability distribution of peak-shaving demand, and thus struggling to support modern power systems' dispatch decisions based on risk warning and probabilistic optimization. Secondly, the input benchmarks of the calculation methods suffer from systematic errors. Key parameters such as the minimum technical output of conventional power sources and the planned output of tie lines are typically determined by simple proportional estimation based on historical and future planned values, failing to fully consider the impact of complex operating scenarios such as real-time unit combination and network congestion, resulting in inherent biases at the source of the calculation. Finally, the calculation methods are inefficient. Existing research and practice, with the few methods attempting to consider uncertainty, generally rely on traditional methods such as Monte Carlo simulations when merging multi-source probability distributions such as load and renewable energy output. The enormous computational and time complexity of these methods makes it difficult to meet the timeliness requirements of online rolling optimization and real-time dispatch.

[0013] To at least solve the above problems, please refer to Figure 1 This invention provides a regional-level renewable energy peak-shaving demand analysis method, including the following steps: Based on the obtained input feature vectors reflecting the operating conditions of the power system, the correction coefficients for the minimum technical output of conventional power sources and the switching power of tie lines are obtained using the trained base model pool. Obtain planned values ​​for the minimum technical output of conventional power supplies and the switching power of tie lines; The planned value is corrected using the correction factor to obtain the corrected minimum technical output of conventional power and the corrected tie-line switching power. Multiple training datasets were constructed using a block resampling strategy based on historical electricity load time-series data and historical regional total renewable energy output time-series data. Train an extreme learning machine model using each training subset in the training dataset to obtain multiple trained extreme learning machine models. The total prediction error of the multiple trained Extreme Learning Machine models is obtained based on the cumulative effect in statistics. n The order of cumulative quantity, and based on the total prediction error. n The cumulative quantities are expanded using Gram-Charlier A series to obtain the first probability density function of the electricity load and the second probability density function of the total output of regional new energy sources. The peak demand probability distribution is calculated using a Gaussian mixture model based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the first probability density function, and the second probability density function.

[0014] As described above, the beneficial effects of this invention are as follows: Based on the obtained input feature vector reflecting the operating conditions of the power system, correction coefficients for the minimum technical output of conventional power sources and the switching power of tie lines are obtained using a trained base model pool. These correction coefficients are then used to correct the planned values ​​of the minimum technical output of conventional power sources and the switching power of tie lines. An ensemble learning strategy is introduced, effectively correcting the systematic deviations in the planned values ​​of the minimum technical output of conventional power sources and the switching power of tie lines. Next, multiple training datasets are constructed using a block resampling strategy based on historical electricity load time-series data and historical regional total renewable energy output time-series data. Each training subset is used to train an extreme learning machine model. The total prediction error of the multiple trained extreme learning machine models is obtained based on the cumulative effect in statistics. n The order cumulative, and based on the total prediction error n The cumulative quantities are expanded using Gram-Charlier A series to obtain the first probability density function of the electricity load and the second probability density function of the total output of regional renewable energy. This preserves the temporal fluctuation characteristics of renewable energy output. By synthesizing cumulative quantities, data and model uncertainties are captured, overcoming the limitations of traditional methods that are based on assumptions of a specific probability distribution. Finally, based on the modified minimum technical output of conventional power sources, the modified tie-line switching power, the first probability density function, and the second probability density function, a Gaussian mixture model is used to calculate the probability distribution of peak-shaving demand. The use of a Gaussian mixture model to replace the traditional Monte Carlo simulation transforms the complex probability fusion process into efficient matrix algebra operations, significantly reducing computational and time complexity, thereby improving the accuracy and efficiency of peak-shaving demand analysis.

[0015] Furthermore, before obtaining the correction coefficients for the minimum technical output of conventional power sources and the switching power of tie lines using the trained base model pool based on the obtained input feature vectors reflecting the operating conditions of the power system, the following steps are included: Collect historical data on actual electricity load, actual total output of regional renewable energy, total horizontal irradiance, ambient temperature, wind speed, planned minimum technical output of conventional power sources, planned power of tie line switching, actual minimum technical output of conventional power sources, and actual power of tie line switching. The first error coefficient is calculated based on the actual value of the minimum technical output of the conventional power supply and the planned value of the minimum technical output of the conventional power supply. The second error coefficient is calculated based on the actual value of the tie-line switching power and the planned value of the tie-line switching power. An input feature vector for supervised learning is established based on the actual values ​​of electricity load, total output of regional new energy sources, total horizontal irradiance, ambient temperature, wind speed, planned minimum technical output of conventional power sources, planned value of tie-line switching power, actual value of minimum technical output of conventional power sources, and actual value of tie-line switching power at the historical time. A sample set for supervised learning is established based on the input feature vector, the first error coefficient, and the second error coefficient. A base model pool is constructed based on the random forest model, the lightweight gradient booster model, and the multilayer perceptron. The base model pool is trained using the sample set used for supervised learning to obtain the trained base model pool.

[0016] As described above, the operating state of a power system is significantly affected by multiple factors such as weather, load behavior, and dispatch plans. A single-dimensional variable cannot fully characterize the system's features. By collecting actual electricity load values ​​and the actual total output of regional renewable energy sources, the temporal correlation between user electricity consumption habits and renewable energy output can be described. Collecting total horizontal irradiance, ambient temperature, and wind speed as environmental variables, and collecting planned values ​​of minimum technical output from conventional power sources and planned values ​​of tie-line switching power, can describe the power grid's human intervention and planned status, and define the system's output benchmark. The first error coefficient is calculated based on the actual and planned values ​​of minimum technical output from conventional power sources. The second error coefficient is calculated by comparing the actual and planned values ​​of the tie-line exchange power. This coefficient describes the deviation characteristics of historical data, facilitating the training of the correction model. Establishing an input feature vector for supervised learning allows for the concatenation of multi-source discrete physical parameters, forming a standard input interface for machine learning, deep learning, and other models. This fully defines the characteristic operating scenarios of the power system at each moment in the mathematical space. By establishing a sample set for supervised learning, the physical problems in this power system domain can be mapped to mathematical regression problems. By building a base model pool based on the random forest model, lightweight gradient booster model, and multilayer perceptron model, and then training it, a high-precision correction model can be obtained.

[0017] Furthermore, based on the obtained input feature vectors reflecting the operating conditions of the power system, the correction coefficients for the minimum technical output of conventional power sources and the switching power of tie lines are obtained using the trained base model pool, including: The obtained input feature vector reflecting the operating conditions of the power system is input into the trained base model pool, and the first initial correction coefficient for the minimum technical output of the conventional power source and the second initial correction coefficient for the tie-line switching power are output. Calculate the Euclidean distance between the input feature vector and historical samples in the preset historical database; Select the closest distance based on the Euclidean distance. K Historical samples from each historical moment constitute a set of nearest neighbors; Calculate the root mean square error of each base model in the trained base model pool on the nearest neighbor set; Based on the root mean square error, the reciprocal method is used to calculate the dynamic weights to obtain the output weights of each base model. The output weights are weighted and combined with the first initial correction coefficient and the second initial correction coefficient respectively to obtain the first correction coefficient for the minimum technical output of the conventional power supply and the second correction coefficient for the switching power of the tie line.

[0018] As described above, the initial correction coefficients are output from the trained base model pool. The historical performance of the algorithm model in similar scenarios can indirectly evaluate the reliability of its current prediction. Therefore, based on the initial correction coefficients, a dynamic weight allocation mechanism based on the K-nearest neighbor method is set up. The nearest neighbor set is formed by searching according to the scenario similarity, and then the root mean square error is calculated. This index can reflect the predictive ability of the base model under the current type of working condition. Then, the reciprocal method is used to calculate the dynamic weights. The smaller the error, the larger the output weight of the model. The output weights are weighted and combined with the first initial correction coefficient and the second initial correction coefficient respectively to obtain the final correction coefficients, which further improves the reliability of the minimum technical output of conventional power supplies and the switching power of tie lines.

[0019] Furthermore, based on historical electricity load time-series data and historical regional total renewable energy output time-series data, multiple training datasets were constructed using a block resampling strategy, including: Historical electricity load time series data and historical regional total renewable energy output time series data are divided into continuous data blocks of preset lengths. The training dataset is obtained by resampling the continuous data blocks. The process of dividing historical electricity load time-series data and historical regional total new energy output time-series data into continuous data blocks of preset lengths is repeated multiple times to obtain multiple training datasets.

[0020] As described above, in order to address the potential defects of traditional methods in destroying the autocorrelation of time series, this invention adopts a block resampling strategy to obtain multiple training datasets. This fully preserves the regional renewable energy power fluctuation patterns such as the continuity and anti-peak characteristics of wind power, the diurnal characteristics of photovoltaic power output, and the load continuity fluctuation characteristics caused by multiple factors, so that the trained model can better capture the temporal dependence of the predicted object.

[0021] Furthermore, the total prediction error of the multiple trained extreme learning machine models is obtained based on the cumulative effect in statistics. n The order cumulants include: The model prediction values ​​are obtained based on the degree of dispersion of the predicted values ​​from the multiple trained extreme learning machine models. n Accumulative quantity; The model data noise is obtained based on the residual distribution of the multiple trained Extreme Learning Machine models on the validation set. n Accumulative quantity; The model's predicted values n The order cumulant and the noise of the model data n The cumulative values ​​of each order are summed to obtain the total prediction error of the multiple trained extreme learning machine models. n Accumulated quantity.

[0022] As described above, based on the statistical concept of cumulants, the uncertainty in the training computation process can be divided into two parts: cognitive uncertainty and random uncertainty. Cognitive uncertainty stems from insufficient training data, manifested as the dispersion of the predicted values ​​of multiple base models. Random uncertainty mainly originates from noise in the data itself. This part of the uncertainty is determined by the distribution of the model's residuals (i.e., the deviation between the true and predicted values) on the validation set. Utilizing the additivity principle of cumulants, the total prediction error... n The cumulants can be obtained from the model predictions. n cumulants and model data noise n The result is obtained by adding the cumulative quantities of each step.

[0023] Furthermore, the calculation of the peak-shaving demand probability distribution using a Gaussian mixture model based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the first probability density function, and the second probability density function includes: The first probability density function and the second probability density function are fitted using a Gaussian mixture model to obtain the fitted first probability density function and the fitted second probability density function. The peak demand probability distribution is obtained using a Gaussian mixture model based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the fitted first probability density function, and the fitted second probability density function.

[0024] Furthermore, the fitted first probability density function is: ; In the formula, This represents the first probability density function after fitting. K This represents the number of Gaussian components in the first probability density function. The first probability density function represents the... k The weights of the Gaussian components, N () represents the Gaussian probability distribution function. The first probability density function represents the... k The mean of the Gaussian components, The first probability density function represents the... k The variance of each Gaussian component; The fitted second probability density function is: ; In the formula, This represents the second probability density function after fitting. J This represents the number of Gaussian components in the second probability density function. The second probability density function represents the first... j The weights of the Gaussian components, The second probability density function represents the first... j The mean of the Gaussian components, The second probability density function represents the first... j The variance of each Gaussian component.

[0025] As described above, the mathematical forms of the probability results of regional renewable energy output and load power obtained in each training session may not be the same. The first probability density function and the second probability density function cannot be directly used for subsequent addition and subtraction operations. Therefore, a Gaussian mixture model is introduced to normalize the random variables (probability distribution forms) required for peak demand calculation. The fitted first probability density function and the fitted second probability density function can then participate in subsequent calculations, thereby improving the efficiency of peak demand analysis.

[0026] Furthermore, the peak-shaving demand probability distribution is obtained using a Gaussian mixture model based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the fitted first probability density function, and the fitted second probability density function, including: The weights of the new Gaussian components in the peak demand probability distribution are calculated based on the weights of the Gaussian components in the fitted first probability density function and the fitted second probability density function. The mean of the new Gaussian component in the peak demand probability distribution is calculated based on the mean of the Gaussian components in the fitted first probability density function and the fitted second probability density function, the corrected minimum technical output of the conventional power supply, and the corrected tie-line switching power. The variance of the new Gaussian component in the peak-shaving demand probability distribution is calculated based on the variance of the Gaussian component in the fitted first probability density function and the fitted second probability density function. The peak-shaving demand probability distribution is obtained based on the weights, mean, and variance of the new Gaussian components after the combination.

[0027] Furthermore, the weights of the new Gaussian components after combination in the peak-shaving demand probability distribution are: ; In the formula, This represents the weight of the new Gaussian component after combination in the probability distribution of peak-shaving demand; The mean of the new Gaussian component after combination in the peak-shaving demand probability distribution is: ; In the formula, This represents the mean of the new Gaussian components after combination in the probability distribution of peak-shaving demand. This indicates the minimum technical output of the modified conventional power supply. This indicates the corrected tie-line switching power; The variance of the new Gaussian component after combination in the peak-shaving demand probability distribution is: ; In the formula, This represents the variance of the new Gaussian component after combination in the probability distribution of peak-shaving demand. This represents the correlation coefficient between regional load and renewable energy output. The first probability density function represents the... k The standard deviation of each Gaussian component The second probability density function represents the first... j The standard deviation of each Gaussian component; The probability distribution of peak-shaving demand is as follows: ; In the formula, This represents the probability distribution of peak-shaving demand.

[0028] As described above, traditional analytical methods often use Monte Carlo simulation (MCS), but this usually requires tens of thousands of random samplings, which is time-consuming and the results are difficult to have universal significance. Therefore, by utilizing the analytical properties of Gaussian mixture models, zero-sampling analytical calculations are achieved. The calculation process is entirely based on matrix algebra operations, which has low complexity and is more efficient.

[0029] Please refer to Figure 2 Another embodiment of the present invention provides a regional-level renewable energy peak-shaving demand analysis system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various steps of the aforementioned regional-level renewable energy peak-shaving demand analysis method.

[0030] The regional-level renewable energy peak-shaving demand analysis method and system described above are applicable to regional-level renewable energy peak-shaving demand analysis scenarios. The specific implementation methods are described below: Please refer to Figure 1 One embodiment of the present invention is as follows: A regional-level renewable energy peak-shaving demand analysis method includes the following steps: Regional renewable energy peak shaving demand D t The computational model is defined as follows: ; In the formula, t Indicates the predicted time. Indicates electrical load. This indicates the region's total renewable energy output, including offshore wind power and solar power. Among them, and For random variables, and For deterministic variables, It reflects the benchmark capability of conventional power supplies to smooth out fluctuations.

[0031] S1. Collect historical data including actual electricity load, total regional renewable energy output, total horizontal irradiance, ambient temperature, wind speed, planned minimum technical output of conventional power sources, planned power of tie line switching, actual minimum technical output of conventional power sources, and actual power of tie line switching.

[0032] Among them, the actual value of electricity load and the actual value of total regional renewable energy output are state variables, the total horizontal irradiance, ambient temperature and wind speed are environmental variables, and the planned value of minimum technical output of conventional power sources and the planned value of tie line switching power are system planned variables.

[0033] In one optional implementation, start-up and shutdown state vectors (including start-up, shutdown, power limiting, etc.) of conventional power sources and new energy units can also be collected, which can affect the minimum technical output of the power source and the prediction results of new energy sources.

[0034] In one alternative implementation, the aforementioned data is collected through channels such as SCADA (Supervisory Control and Data Acquisition), WAMS (Wide Area Measurement System), and meteorological service interfaces.

[0035] S2. Calculate the first error coefficient based on the actual value of the minimum technical output of the conventional power supply and the planned value of the minimum technical output of the conventional power supply, specifically as follows: ; In the formula, This represents the first error coefficient. This represents the actual value of the minimum technical output of a conventional power supply. This represents the planned minimum technical output of a conventional power source.

[0036] S3. Calculate the second error coefficient based on the actual value of the tie-line switching power and the planned value of the tie-line switching power, specifically as follows: ; In the formula, This represents the second error coefficient. This represents the actual value of the switching power of the tie line. This indicates the planned value for the switching power of the tie line.

[0037] Among them, the first error coefficient and the second error coefficient are error observation variables.

[0038] S4. Based on the actual values ​​of electricity load, total regional renewable energy output, total horizontal irradiance, ambient temperature, wind speed, planned minimum technical output of conventional power sources, planned tie-line switching power, actual minimum technical output of conventional power sources, and actual tie-line switching power at the historical time, establish an input feature vector for supervised learning, specifically: ; In the formula, This represents the input feature vector used for supervised learning. express The actual electrical load at any given time. express The actual electrical load at any given time. express Actual value of total regional renewable energy output at any given time. express Actual value of total regional renewable energy output at any given time. Indicates a time step. express t Total horizontal irradiance at time 10:00 express t The ambient temperature at any given time express t The wind speed at any given moment.

[0039] S5. Establish a sample set for supervised learning based on the input feature vector used for supervised learning, the first error coefficient, and the second error coefficient, specifically as follows: ; In the formula, This represents the sample set used for supervised learning. H This represents a set of historical moments. By establishing a sample set, physical problems in this power system field can be mapped to mathematical regression problems.

[0040] In traditional demand calculations, the planned values ​​for minimum technical output of conventional power sources and tie-line switching power are often based on experience or simple scaling rules. However, in actual operation, due to factors such as unit status and the grid at the other end, there is usually a significant systematic deviation between the actual and planned values. Traditional fixed correction methods (such as scaling by the ratio of installed capacity before and after and multiplying by a fixed coefficient) cannot adapt to the changing operating modes of the power grid. Therefore, this invention considers adopting an ensemble learning strategy based on dynamic neighborhood performance evaluation. This strategy dynamically calculates the combined weights based on the performance of each base model in historical scenarios with similar current input features, rather than training a fixed meta-model, thereby achieving adaptive correction for different operating conditions.

[0041] To avoid bias in certain aspects of a single model's performance, a multi-heterogeneous model ensemble strategy is adopted, as detailed below.

[0042] S6. Construct a base model pool based on the random forest model, lightweight gradient booster model, and multilayer perceptron model.

[0043] Among them, the Random Forest model excels at capturing nonlinear relationships of high-dimensional features and has strong resistance to overfitting; the Lightweight Gradient Boosting Machine model has high training efficiency and is sensitive to imbalanced data; and the Multilayer Perceptron model has powerful function approximation capabilities and is good at mining deep features.

[0044] S7. The base model pool is trained using the sample set used for supervised learning to obtain the trained base model pool.

[0045] Specifically, the input feature vector used for supervised learning is used as the model input feature, and the first error coefficient and the second error coefficient are used as the model output label. Each base model in the base model pool is trained to obtain the trained base model pool.

[0046] S8. Based on the obtained input feature vectors reflecting the operating conditions of the power system, the trained base model pool is used to obtain correction coefficients for the minimum technical output of conventional power sources and the switching power of tie lines, specifically including S8.1-S8.6: S81. Input the obtained input feature vector reflecting the operating conditions of the power system into the trained base model pool, and output the first initial correction coefficient of the minimum technical output of the conventional power source and the second initial correction coefficient of the tie-line switching power.

[0047] The initial correction factor is: ; In the formula, Indicates the first m The initial correction coefficients output by each trained base model. Indicates the first m The parameters of a trained base model Indicates the first m A trained base model.

[0048] S8.2 Calculate the Euclidean distance between the input feature vector and historical samples in the preset historical database, specifically as follows: ; In the formula, d Represents the input feature vector Compared with historical samples The Euclidean distance.

[0049] S8.3 Select the closest distance according to the Euclidean distance. KHistorical samples from each historical moment constitute a set of nearest neighbors. N new .

[0050] S8.4 Calculate the root mean square error of each base model in the trained base model pool on the nearest neighbor set, specifically as follows: ; In the formula, Indicates the first m The trained base models are in the nearest neighborhood set N new The root mean square error on Indicates the first j Historical samples (i.e., the nearest neighbor set) N new The first in j The actual error coefficient value corresponding to each sample.

[0051] S8.5. Based on the root mean square error, the reciprocal method is used to calculate the dynamic weights to obtain the output weights of each base model, specifically: ; In the formula, Indicates the first m The output weights of the trained base model Indicates the first i The trained base models are in the nearest neighborhood set N new The root mean square error on This represents the adjustment factor (usually 1 or 2), used to amplify the performance differences between models.

[0052] S8.6. The output weights are weighted and combined with the first initial correction coefficient and the second initial correction coefficient respectively to obtain the first correction coefficient for the minimum technical output of the conventional power supply and the second correction coefficient for the switching power of the tie line.

[0053] The correction factor is: ; In the formula, This represents the correction factor.

[0054] S9. Obtain the planned values ​​for the minimum technical output of the conventional power supply and the switching power of the tie line.

[0055] S10. The planned value is corrected using the correction coefficient to obtain the corrected minimum technical output of the conventional power supply and the corrected tie-line switching power.

[0056] The revised minimum technical output of conventional power supplies is as follows: .

[0057] The output and load of new energy sources such as offshore wind power and distributed photovoltaics are highly stochastic and generally do not follow a standard normal distribution. Traditional deterministic modeling and prediction methods cannot fully quantify the absorption risk of the system. To address this, this invention employs an improved Combined Bootstrap and Cumulants (CBC) algorithm, aiming to simultaneously quantify data uncertainty and model cognitive uncertainty, and to provide highly reliable and complete probability information through analytical methods.

[0058] S11. Based on historical electricity load time-series data and historical regional total renewable energy output time-series data, multiple training datasets are constructed using a block resampling strategy, specifically including S11.1-S11.3: To address the potential drawback of traditional methods in destroying the autocorrelation of time series data, this invention designs a block resampling strategy, as detailed below.

[0059] S11.1 Divide the historical electricity load time series data and the historical regional total renewable energy output time series data into preset lengths respectively. L block Continuous data blocks.

[0060] For example, a data block consists of 8 consecutive hours or 32 points.

[0061] S11.2. Resample the continuous data blocks to obtain the training dataset.

[0062] Specifically, the training dataset is obtained by extracting data with replacement from consecutively divided data blocks.

[0063] S11.3, Return to execute S11.1-S11.2 multiple times to obtain multiple training datasets.

[0064] For example, return to execute S11.1-S11.2 B Once, it can be generated B A separate training dataset D (1) ,..., D (B) .

[0065] S12. Train an Extreme Learning Machine model using each training subset in the training dataset to obtain multiple trained Extreme Learning Machine models.

[0066] The trained Extreme Learning Machine (ELM) model is based on a feedforward neural network. It randomly generates the weights of the hidden layer nodes and does not require updating. Its output weights can be calculated. (b)The analytical solution is as follows: ; In the formula, This represents the hidden layer output matrix. Represents the target value vector. C This represents the regularization coefficient (also known as the penalty coefficient).

[0067] By using multiple trained ELM models, one can obtain B A parallel predictor, resulting in B Different judgments and outcomes are used to more thoroughly characterize and address the cognitive uncertainty of the model.

[0068] S13. The total prediction error of the multiple trained extreme learning machine models is obtained based on the cumulative effect in statistics. n The order of cumulative quantity, and based on the total prediction error. n The cumulative quantities are expanded using Gram-Charlier A series to obtain the first probability density function of the electricity load and the second probability density function of the total output of regional renewable energy, specifically including S13.1-S13.4: S13.1 Obtain the model prediction value based on the degree of dispersion of the predicted values ​​of the multiple trained extreme learning machine models. n Accumulated quantity.

[0069] Among them, the model predicted values n Second-order cumulant (variance) in first-order cumulants For example, it can be represented as: ; In the formula, Indicates the first b The predicted output value of an extreme learning machine model. Indicates all B The arithmetic mean of the predicted output values ​​of the extreme learning machine models. .

[0070] S13.2. Obtain the model data noise based on the residual distribution of the multiple trained extreme learning machine models on the validation set. n Step cumulative .

[0071] S13.3, The model predicts the values n The order cumulant and the noise of the model data n The cumulative values ​​of each order are summed to obtain the total prediction error of the multiple trained extreme learning machine models. n Step cumulative Specifically: .

[0072] S13.4, Based on the total prediction error n The cumulative quantities are expanded using Gram-Charlier A series to obtain the first probability density function of the electricity load and the second probability density function of the total output of regional renewable energy.

[0073] Specifically, based on the total prediction error n The third-order total cumulant (correlated skewness) and the fourth-order total cumulant (correlated kurtosis) in the total cumulant are expanded using Gram-Charlier A series to obtain the first probability density function of the electricity load and the second probability density function of the total output of regional renewable energy.

[0074] Wherein, the second probability density function is: ; In the formula, This represents the second probability density function. This represents the probability density function of the standard normal distribution. This represents the total cumulative amount of the third order. This represents the total cumulative amount of the fourth order. Representing Hermitian polynomials, including and .

[0075] The probability density function of electricity load and total regional renewable energy output uses higher-order statistics to correct the normal distribution, accurately describing the asymmetry (skewness) and thick-tailed characteristics (kurtosis) of regional renewable energy output and load power distribution.

[0076] S14. Based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the first probability density function, and the second probability density function, a Gaussian mixture model is used to calculate the probability distribution of peak-shaving demand, specifically including S14.1-S14.2: S14.1. Use a Gaussian mixture model to fit the first probability density function and the second probability density function respectively to obtain the fitted first probability density function and the fitted second probability density function.

[0077] Specifically, based on the Expectation Maximization (EM) algorithm, the first probability density function is fitted as follows: K The weighted sum of the Gaussian components yields the first probability density function after fitting: ; In the formula, This represents the first probability density function after fitting. KThis represents the number of Gaussian components in the first probability density function. The first probability density function represents the... k The weights of the Gaussian components, and satisfying , N () represents the Gaussian probability distribution function. The first probability density function represents the... k The mean of the Gaussian components, The first probability density function represents the... k The variance of each Gaussian component; Based on the expectation-maximization algorithm, the first probability density function is fitted as follows: J The weighted sum of the Gaussian components yields the fitted second probability density function as follows: ; In the formula, This represents the second probability density function after fitting. J This represents the number of Gaussian components in the second probability density function. The second probability density function represents the first... j The weights of the Gaussian components, and satisfying , The second probability density function represents the first... j The mean of the Gaussian components, The second probability density function represents the first... j The variance of each Gaussian component.

[0078] S14.2. Based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the fitted first probability density function, and the fitted second probability density function, a Gaussian mixture model is used to obtain the peak-shaving demand probability distribution, specifically including S14.2.1-S14.2.4: Peak-shaving demand calculation involves random variables ( , ) and deterministic variables ( , This invention utilizes the analytical properties of Gaussian Mixture Models (GMMs) to achieve zero-sampling analytical computation. Traditional analytical methods often use Monte Carlo simulation (MCS), but this typically requires tens of thousands of random samplings, is computationally time-consuming, and the results are rarely universally applicable. Therefore, this invention leverages the analytical properties of GMMs to achieve zero-sampling analytical computation.

[0079] According to probability theory, a linear combination of two independent GMM distributions is still a GMM distribution. This defines peak-shaving demand. probability distribution Depend on K J It consists of a new Gaussian component, as detailed below.

[0080] S14.2.1. The weights of the combined new Gaussian components in the peak-shaving demand probability distribution are calculated based on the weights of the Gaussian components in the fitted first probability density function and the fitted second probability density function, specifically as follows: ; In the formula, This represents the weight of the new Gaussian component after combination in the probability distribution of peak-shaving demand.

[0081] S14.2.2. Based on the mean of the Gaussian components in the fitted first probability density function and the fitted second probability density function, the mean of the new Gaussian component in the peak-shaving demand probability distribution after combination is calculated using the corrected minimum technical output of the conventional power supply and the corrected tie-line switching power. Specifically: ; In the formula, This represents the mean of the new Gaussian components after combination in the probability distribution of peak-shaving demand. This indicates the minimum technical output of the modified conventional power supply. This indicates the corrected switching power of the tie line.

[0082] S14.2.3. The variance of the new Gaussian component in the peak-shaving demand probability distribution is calculated based on the variances of the Gaussian components in the fitted first probability density function and the fitted second probability density function, specifically as follows: ; In the formula, This represents the variance of the new Gaussian component after combination in the probability distribution of peak-shaving demand. This represents the correlation coefficient between regional load and renewable energy output. The first probability density function represents the... k The standard deviation of each Gaussian component The second probability density function represents the first... j The standard deviation of each Gaussian component.

[0083] S14.2.4. Based on the weights, mean, and variance of the combined new Gaussian components, the peak-shaving demand probability distribution is obtained, specifically as follows: ; In the formula, This represents the probability distribution of peak-shaving demand.

[0084] In one alternative implementation, it further includes: The cumulative distribution function of peak-shaving demand is obtained by integrating the probability distribution of peak-shaving demand. ; The existing peak shaving gap probability and confidence peak shaving demand are obtained based on the cumulative distribution function of the peak shaving demand.

[0085] The existing peak-shaving gap probability is: ; The confidence peak-shaving requirement is as follows: ; In the formula, This indicates confidence in peak-shaving demand. Indicates the confidence level.

[0086] This is because the integral of the GMM is in analytical form, allowing for the rapid calculation of quantiles under any probability. Utilizing the probability distribution of peak-shaving demand can provide scheduling agencies with an intuitive basis for decision-making. The existing peak-shaving gap probability reflects the probability that the system's existing conventional power sources cannot meet peak-shaving demands, and the confidence level in peak-shaving demand quantifies the confidence level. The average peak demand at any given time can effectively quantify the reserve capacity required by the power grid in special circumstances such as extreme weather or load surges.

[0087] According to another aspect of the invention, Figure 2 This is a schematic diagram illustrating a regional-level renewable energy peak-shaving demand analysis system according to an embodiment of the present invention. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various steps of the regional-level renewable energy peak-shaving demand analysis method described above.

[0088] In summary, the regional-level renewable energy peak-shaving demand analysis method and system described above, based on the obtained input feature vector reflecting the power system's operating conditions, uses a trained base model pool to obtain correction coefficients for the minimum technical output of conventional power sources and the tie-line exchange power. These correction coefficients are then used to adjust the planned values ​​of the minimum technical output of conventional power sources and the tie-line exchange power. An ensemble learning strategy is introduced to effectively correct systematic deviations in the planned values ​​of the minimum technical output of conventional power sources and the tie-line exchange power. Next, based on historical electricity load time-series data and historical regional renewable energy total output time-series data, a block resampling strategy is used to construct multiple training datasets. Each training subset is used to train an extreme learning machine model. The total prediction error of the multiple trained extreme learning machine models is obtained based on statistical cumulative factors. n The order cumulative, and based on the total prediction error nThe cumulants are expanded using Gram-Charlier A series to obtain the first probability density function of the electricity load and the second probability density function of the total regional renewable energy output. This preserves the time-series fluctuation characteristics of renewable energy output. By synthesizing cumulants, data and model uncertainties are captured, overcoming the limitations of traditional methods that rely on assumptions about specific probability distributions. Finally, based on the corrected minimum technical output of conventional power sources, the corrected tie-line switching power, and the first and second probability density functions, a Gaussian mixture model is used to calculate the probability distribution of peak-shaving demand. Using a Gaussian mixture model instead of traditional Monte Carlo simulation transforms the complex probability fusion process into efficient matrix algebra operations, significantly reducing computational complexity. This reduces time complexity, thereby improving the accuracy and efficiency of peak demand analysis. Furthermore, based on the initial correction coefficients, a dynamic weight allocation mechanism based on the K-nearest neighbor method is established. A nearest neighbor set is formed by searching based on scenario similarity, and then the root mean square error is calculated. This index reflects the predictive ability of the base model under the current operating conditions. Next, the reciprocal method is used for dynamic weight calculation; the smaller the error, the larger the output weight of the model. The output weights are then weighted and combined with the first and second initial correction coefficients to obtain the final correction coefficients, further improving the reliability of the minimum technical output of conventional power supplies and the switching power of tie lines.

[0089] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A regional-level renewable energy peak-shaving demand analysis method, characterized in that, Including the following steps: Based on the obtained input feature vectors reflecting the operating conditions of the power system, the correction coefficients for the minimum technical output of conventional power sources and the switching power of tie lines are obtained using the trained base model pool. Obtain planned values ​​for the minimum technical output of conventional power supplies and the switching power of tie lines; The planned value is corrected using the correction factor to obtain the corrected minimum technical output of conventional power and the corrected tie-line switching power. Multiple training datasets were constructed using a block resampling strategy based on historical electricity load time-series data and historical regional total renewable energy output time-series data. Train an extreme learning machine model using each training subset in the training dataset to obtain multiple trained extreme learning machine models. The total prediction error of the multiple trained Extreme Learning Machine models is obtained based on the cumulative effect in statistics. n The order of cumulative quantity, and based on the total prediction error. n The cumulative quantities are expanded using Gram-Charlier A series to obtain the first probability density function of the electricity load and the second probability density function of the total output of regional new energy sources. The peak demand probability distribution is calculated using a Gaussian mixture model based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the first probability density function, and the second probability density function.

2. The regional-level renewable energy peak-shaving demand analysis method according to claim 1, characterized in that, Before obtaining the correction coefficients for the minimum technical output of conventional power sources and tie-line switching power based on the obtained input feature vectors reflecting the operating conditions of the power system using the trained base model pool, the process includes: Collect historical data on actual electricity load, actual total output of regional renewable energy, total horizontal irradiance, ambient temperature, wind speed, planned minimum technical output of conventional power sources, planned power of tie line switching, actual minimum technical output of conventional power sources, and actual power of tie line switching. The first error coefficient is calculated based on the actual value of the minimum technical output of the conventional power supply and the planned value of the minimum technical output of the conventional power supply. The second error coefficient is calculated based on the actual value of the tie-line switching power and the planned value of the tie-line switching power. An input feature vector for supervised learning is established based on the actual values ​​of electricity load, total output of regional new energy sources, total horizontal irradiance, ambient temperature, wind speed, planned minimum technical output of conventional power sources, planned value of tie-line switching power, actual value of minimum technical output of conventional power sources, and actual value of tie-line switching power at the historical time. A sample set for supervised learning is established based on the input feature vector, the first error coefficient, and the second error coefficient. A base model pool is constructed based on the random forest model, the lightweight gradient booster model, and the multilayer perceptron. The base model pool is trained using the sample set used for supervised learning to obtain the trained base model pool.

3. The regional-level renewable energy peak-shaving demand analysis method according to claim 1, characterized in that, Based on the obtained input feature vectors reflecting the operating conditions of the power system, and using the trained base model pool, correction coefficients for the minimum technical output of conventional power sources and tie-line switching power are obtained, including: The obtained input feature vector reflecting the operating conditions of the power system is input into the trained base model pool, and the first initial correction coefficient for the minimum technical output of the conventional power source and the second initial correction coefficient for the tie-line switching power are output. Calculate the Euclidean distance between the input feature vector and historical samples in the preset historical database; Select the closest distance based on the Euclidean distance. K Historical samples from each historical moment constitute a set of nearest neighbors; Calculate the root mean square error of each base model in the trained base model pool on the nearest neighbor set; Based on the root mean square error, the reciprocal method is used to calculate the dynamic weights to obtain the output weights of each base model. The output weights are weighted and combined with the first initial correction coefficient and the second initial correction coefficient to obtain the first correction coefficient for the minimum technical output of the conventional power supply and the second correction coefficient for the switching power of the tie line.

4. The regional-level renewable energy peak-shaving demand analysis method according to claim 1, characterized in that, Multiple training datasets were constructed using a block resampling strategy based on historical electricity load time-series data and historical regional total renewable energy output time-series data, including: Historical electricity load time series data and historical regional total renewable energy output time series data are divided into continuous data blocks of preset lengths. The training dataset is obtained by resampling the continuous data blocks. The process of dividing historical electricity load time-series data and historical regional total new energy output time-series data into continuous data blocks of preset lengths is repeated multiple times to obtain multiple training datasets.

5. The regional-level renewable energy peak-shaving demand analysis method according to claim 1, characterized in that, The total prediction error of the multiple trained Extreme Learning Machine models is obtained based on the cumulative effect in statistics. n The order cumulants include: The model prediction values ​​are obtained based on the degree of dispersion of the predicted values ​​from the multiple trained extreme learning machine models. n Accumulative quantity; The model data noise is obtained based on the residual distribution of the multiple trained Extreme Learning Machine models on the validation set. n Accumulative quantity; The model's predicted values n The order cumulant and the noise of the model data n The cumulative values ​​of each order are summed to obtain the total prediction error of the multiple trained extreme learning machine models. n Accumulated quantity.

6. The regional-level renewable energy peak-shaving demand analysis method according to claim 1, characterized in that, The peak-shaving demand probability distribution is calculated using a Gaussian mixture model based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the first probability density function, and the second probability density function, including: The first probability density function and the second probability density function are fitted using a Gaussian mixture model to obtain the fitted first probability density function and the fitted second probability density function. The peak demand probability distribution is obtained using a Gaussian mixture model based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the fitted first probability density function, and the fitted second probability density function.

7. The regional-level renewable energy peak-shaving demand analysis method according to claim 6, characterized in that, The fitted first probability density function is: ； In the formula, This represents the first probability density function after fitting. K This represents the number of Gaussian components in the first probability density function. The first probability density function represents the... k The weights of the Gaussian components, N () represents the Gaussian probability distribution function. The first probability density function represents the... k The mean of the Gaussian components, The first probability density function represents the... k The variance of each Gaussian component; The fitted second probability density function is: ； In the formula, This represents the second probability density function after fitting. J This represents the number of Gaussian components in the second probability density function. The second probability density function represents the first... j The weights of the Gaussian components, The second probability density function represents the first... j The mean of the Gaussian components, The second probability density function represents the first... j The variance of each Gaussian component.

8. The regional-level renewable energy peak-shaving demand analysis method according to claim 7, characterized in that, The peak-shaving demand probability distribution is obtained using a Gaussian mixture model based on the modified minimum technical output of the conventional power supply, the modified tie-line switching power, the fitted first probability density function, and the fitted second probability density function, including: The weights of the new Gaussian components in the peak demand probability distribution are calculated based on the weights of the Gaussian components in the fitted first probability density function and the fitted second probability density function. The mean of the new Gaussian component in the peak demand probability distribution is calculated based on the mean of the Gaussian components in the fitted first probability density function and the fitted second probability density function, the corrected minimum technical output of the conventional power supply, and the corrected tie-line switching power. The variance of the new Gaussian component in the peak-shaving demand probability distribution is calculated based on the variance of the Gaussian component in the fitted first probability density function and the fitted second probability density function. The peak-shaving demand probability distribution is obtained based on the weights, mean, and variance of the new Gaussian components after the combination.

9. The regional-level renewable energy peak-shaving demand analysis method according to claim 8, characterized in that, The weights of the new Gaussian components after combination in the peak-shaving demand probability distribution are: ； In the formula, This represents the weight of the new Gaussian component after combination in the probability distribution of peak-shaving demand; The mean of the new Gaussian component after combination in the peak-shaving demand probability distribution is: ； In the formula, This represents the mean of the new Gaussian components after combination in the probability distribution of peak-shaving demand. This indicates the minimum technical output of the modified conventional power supply. This indicates the corrected tie-line switching power; The variance of the new Gaussian component after combination in the peak-shaving demand probability distribution is: ； In the formula, This represents the variance of the new Gaussian component after combination in the probability distribution of peak-shaving demand. This represents the correlation coefficient between regional load and renewable energy output. The first probability density function represents the... k The standard deviation of each Gaussian component The second probability density function represents the first... j The standard deviation of each Gaussian component; The probability distribution of peak-shaving demand is as follows: ； In the formula, This represents the probability distribution of peak-shaving demand.

10. A regional-level renewable energy peak-shaving demand analysis system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements each step of the regional-level new energy peak-shaving demand analysis method according to any one of claims 1 to 9.

Images