A Multi-Scale Power Prediction and Uncertainty Quantification Method for Photovoltaic-Storage-Charging Microgrids

Abstract

This invention discloses a multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids, belonging to the field of power dispatching technology. The method includes: acquiring historical operational data and meteorological data of the photovoltaic-storage-charging microgrid; extracting different time-frequency features through a multi-scale decomposition algorithm; establishing a nonlinear mapping relationship through a deep neural network prediction model to determine the optimal network structure and hyperparameter combination; quantifying prediction uncertainty using a Bayesian deep learning framework to obtain prediction confidence intervals; training multiple base learners with different structures and obtaining ensemble prediction results through an ensemble learning algorithm; and obtaining a power prediction solution for the photovoltaic-storage-charging microgrid by constructing prediction intervals and combining them with point prediction accuracy indicators. This invention achieves high-precision power prediction and accurate uncertainty quantification for photovoltaic-storage-charging microgrids, providing reliable decision support for the safe and stable operation of distribution networks.

Description

Technical Field

[0001] This invention relates to the field of power dispatching technology, and in particular to a multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids. Background Technology

[0002] With the large-scale application of photovoltaic-storage-charging microgrids in distribution networks, accurate power forecasting is crucial for the safe and stable operation of the system. Traditional power forecasting methods, mainly based on time series analysis and machine learning techniques, have significant limitations in handling the complex multi-scale characteristics of photovoltaic-storage-charging microgrids. Existing technologies suffer from the following main drawbacks: traditional forecasting methods typically employ single-time-scale modeling, failing to effectively capture the multi-time-scale fluctuations of photovoltaic power generation, the periodicity of energy storage system charging and discharging, and the randomness of electric vehicle charging, resulting in limited forecast accuracy. Furthermore, existing methods lack effective quantification of forecast uncertainty, failing to provide risk assessment information for grid dispatch and hindering reliable decision-making.

[0003] While existing deep learning prediction methods possess powerful nonlinear mapping capabilities, they suffer from insufficient feature extraction and poor model interpretability when handling the fusion of multi-source heterogeneous data. This is particularly true in complex systems like photovoltaic-storage-charging microgrids, which contain various equipment types and operating modes, where traditional methods struggle to establish effective feature correlations. Furthermore, existing uncertainty quantification methods are primarily based on statistical assumptions and lack in-depth consideration of physical mechanisms, often resulting in inaccurate uncertainty estimations when facing extreme weather conditions or equipment failures. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids to solve the problems of complex multi-scale characteristics, difficulty in uncertainty quantification, and limited prediction accuracy in existing photovoltaic-storage-charging microgrid power prediction.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids, including:

[0008] Acquire historical operational data and meteorological data for the photovoltaic-storage-charging microgrid;

[0009] Based on the preprocessed data, different time-frequency features are extracted using a multi-scale decomposition algorithm;

[0010] Based on different time-frequency characteristics, a nonlinear mapping relationship is established through a deep neural network prediction model to determine the optimal network structure and hyperparameter combination.

[0011] Based on the optimal network structure and hyperparameter combination, the prediction uncertainty is quantified and the prediction confidence interval is obtained through the Bayesian deep learning framework.

[0012] Based on the predicted confidence interval, multiple base learners with different structures are trained, and the ensemble prediction result is obtained through an ensemble learning algorithm.

[0013] Based on the integrated prediction results, a power prediction solution for photovoltaic-storage-charging microgrids is obtained by constructing prediction intervals and combining them with point prediction accuracy indicators.

[0014] As a preferred embodiment of the multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids described in this invention, the extraction of different time-frequency features includes:

[0015] By performing discrete wavelet decomposition on the total power signal of the photovoltaic-storage-charging microgrid, the wavelet detail coefficients and approximation coefficients of each layer are obtained;

[0016] For photovoltaic power generation, multi-scale components are divided according to physical laws, and decomposed into trend term, daily periodic term and weather disturbance term.

[0017] By performing wavelet decomposition on the original power signal of the photovoltaic-storage-charging microgrid and calculating the sum of squares of the wavelet detail coefficients of each layer, the energy characteristics of each layer are extracted.

[0018] Based on the trend term, daily periodic term, and weather disturbance term, the power spectral density of photovoltaic power is extracted by performing a Fourier transform on the photovoltaic power generation.

[0019] By performing continuous wavelet transform on the total power signal of the system, the time-frequency energy distribution is constructed, and the energy at the same frequency is normalized along the time dimension to obtain the time-frequency correlation characteristics.

[0020] Based on the energy characteristics, power spectral density, and time-frequency correlation characteristics of each layer, different time-frequency features are extracted.

[0021] As a preferred embodiment of the multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids described in this invention, wherein: determining the optimal network structure and hyperparameter combination includes:

[0022] Meteorological, historical power, and calendar features are concatenated into a fusion input vector;

[0023] Based on the fused input vector, a multi-layer nonlinear transformation is defined;

[0024] Based on nonlinear transformation, attention scores are calculated, and attention weights are obtained through normalization.

[0025] Based on the deviation between the attention weights and the true values, the prediction loss is calculated.

[0026] Based on prediction loss, a total loss function is constructed by combining uncertainty loss and regularization loss;

[0027] Based on the total loss function, activation function, optimizer, and learning rate scheduling strategy are set, and the optimal network structure and hyperparameter combination are determined through cross-validation.

[0028] As a preferred embodiment of the multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids described in this invention, wherein: obtaining the prediction confidence interval includes:

[0029] By using a Bayesian deep learning framework, parameter uncertainty is expressed as a posterior distribution;

[0030] Based on the posterior distribution, the prediction mean and total prediction variance are calculated to obtain the prediction uncertainty.

[0031] Define uncertainty loss, take the prediction uncertainty as the learning objective, and obtain the prediction confidence interval.

[0032] As a preferred embodiment of the multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids described in this invention, the method for obtaining integrated prediction results includes:

[0033] Define the weighted average form of the ensemble forecast;

[0034] Based on the weighted average form, the objective function is optimized on the validation set to obtain the optimized weights;

[0035] Based on the optimized weights, the weights are dynamically adjusted according to the real-time root mean square error of each model and the exponential decay law through a dynamic weight update mechanism to obtain the updated weights.

[0036] The integrated prediction results are obtained by normalizing the updated weights.

[0037] As a preferred embodiment of the multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids described in this invention, the calculation of the total prediction variance includes using Bayesian inference to capture model uncertainty, while introducing observation noise terms to model data uncertainty.

[0038] As a preferred embodiment of the multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids described in this invention, the design uncertainty loss includes a variance weighting term of the prediction error and a logarithmic regularization term of the prediction variance.

[0039] As a preferred embodiment of the multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids described in this invention, the dynamic weight update mechanism is expressed as follows:

[0040] ;

[0041] in, For the updated unnormalized weights, The unnormalized weights before the update. The attenuation coefficient is... The root mean square error, It is an exponential function. This refers to the dynamic weight update cycle of the ensemble model.

[0042] As a preferred embodiment of the multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids described in this invention, the time-frequency correlation feature is expressed as:

[0043] ;

[0044] in, It is a time-frequency correlation. To charge the total power of the photovoltaic-storage microgrid, The translation parameter in continuous wavelet transform. For wavelet transform, For integration time variable, Integrating along the time axis, The time variable represents the moment when the power signal changes over time. For frequency variables, it represents the frequency components of the signal in the frequency domain.

[0045] As a preferred embodiment of the multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids described in this invention, wherein: the design uncertainty loss Represented as:

[0046] ;

[0047] in, Let K be the prediction variance of the k-th sample, and K be the total number of samples. As the average factor, For summation, The square of the prediction error. For the standardized error term, For logarithmic variance regularization term, Let be the predicted mean of the k-th sample. This represents the true power value of the k-th sample.

[0048] The beneficial effects of this invention are as follows: This invention establishes for the first time a multi-scale power prediction framework for photovoltaic-storage-charging microgrids. By extracting different time-frequency features through wavelet analysis, it effectively captures the dynamic characteristics of the system across multiple time scales, significantly improving prediction accuracy. It constructs an uncertainty quantification model based on Bayesian deep learning, achieving accurate estimation of prediction uncertainty and providing reliable risk assessment information for power grid dispatch. A multi-model ensemble learning algorithm is developed, improving the robustness and adaptability of prediction through a dynamic weight adjustment mechanism, maintaining excellent performance under different operating scenarios. A complete prediction evaluation index system is established, focusing not only on prediction accuracy but also on the quality of uncertainty assessment, providing a quantitative basis for method improvement. Simulation results show that compared with traditional methods, significant improvements are achieved in key indicators such as prediction accuracy, uncertainty quantification, and computational efficiency. This provides important technical support for the intelligent operation of photovoltaic-storage-charging microgrids and the safe and stable operation of distribution networks, possessing significant theoretical value and broad engineering application prospects. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram of the basic process of a multi-scale power prediction and uncertainty quantification method for a photovoltaic-storage-charging microgrid provided in one embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of the basic process of a multi-scale power prediction and uncertainty quantification method for a photovoltaic-storage-charging microgrid provided in one embodiment of the present invention. Detailed Implementation

[0052] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0053] Example 1, referring to Figure 1 As an embodiment of the present invention, a multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids is provided, comprising:

[0054] S100: Acquire historical operational data and meteorological data of the photovoltaic-storage-charging microgrid;

[0055] In this embodiment of the invention, the historical operating data of the photovoltaic-storage-charging microgrid in step S100 includes photovoltaic power generation, energy storage charging and discharging power, and electric vehicle charging power. Simultaneously, relevant meteorological data, such as solar radiation intensity, ambient temperature, humidity, and wind speed, are collected. The raw data undergoes quality checks to identify and process outliers and missing values. Methods such as sliding window averaging and median filtering are used to reduce noise interference. A data standardization and normalization process is established to ensure the comparability of data at different scales.

[0056] S200: Based on the preprocessed data, different time-frequency features are extracted through a multi-scale decomposition algorithm;

[0057] In this embodiment of the invention, the multi-scale decomposition algorithm in step S200 includes:

[0058] By performing discrete wavelet decomposition on the total power signal of the photovoltaic-storage-charging microgrid, the wavelet detail coefficients and approximation coefficients of each layer are obtained;

[0059] For photovoltaic power generation, multi-scale components are divided according to physical laws, and decomposed into trend term, daily periodic term and weather disturbance term.

[0060] By performing wavelet decomposition on the original power signal of the photovoltaic-storage-charging microgrid and calculating the sum of squares of the wavelet detail coefficients of each layer, the energy characteristics of each layer are extracted.

[0061] Based on the trend term, daily periodic term, and weather disturbance term, the power spectral density of photovoltaic power is extracted by performing a Fourier transform on the photovoltaic power generation.

[0062] By performing continuous wavelet transform on the total power signal of the system, the time-frequency energy distribution is constructed, and the energy at the same frequency is normalized along the time dimension to obtain the time-frequency correlation characteristics.

[0063] Based on the energy characteristics, power spectral density, and time-frequency correlation characteristics of each layer, different time-frequency features are extracted.

[0064] In this embodiment of the invention, the total power signal of the optical-storage-charging microgrid is represented by discrete wavelet decomposition as follows:

[0065] ;

[0066] in, To charge the total power of the photovoltaic-storage microgrid, For scale indexing, This represents the total number of decomposition levels. For translation index, For detail coefficients, For wavelet basis functions, These are approximate coefficients. It is a scaling function.

[0067] In this embodiment of the invention, the photovoltaic power generation is represented by multi-scale component division based on physical laws as follows:

[0068]

[0069] in, Photovoltaic power generation capacity, As a trend component, For daily periodic components, This represents the intraday fluctuation component. This represents the noise component.

[0070] In this embodiment of the invention, the multi-scale decomposition algorithm further includes establishing a state-dependent dynamic model of the energy storage system power:

[0071]

[0072] in, For the power of the energy storage system, For energy storage control strategy functions, In a charged state, To meet load demand, This is an electricity price signal.

[0073] It should be noted that, as an "energy storage state model," although it does not directly participate in feature calculation, it provides a physical mechanism explanation for the dynamic patterns in the total power and is a key link in realizing "mechanism-data fusion."

[0074] In this embodiment of the invention, the extracted energy features of each layer are represented as follows:

[0075] ;

[0076] in, Let be the energy characteristic of the j-th layer.

[0077] In this embodiment of the invention, the power spectral density of the extracted photovoltaic power is expressed as:

[0078]

[0079] in, The power spectral density of photovoltaic power. For Fourier transform operators, For frequency variables.

[0080] In this embodiment of the invention, the time-frequency correlation feature is represented as follows:

[0081] ;

[0082] in, It is a time-frequency correlation. The translation parameter in continuous wavelet transform. For wavelet transform, For integration time variable, Integrating along the time axis, The time variable represents the moment when the power signal changes over time. For frequency variables, it represents the frequency components of the signal in the frequency domain.

[0083] In this embodiment of the invention, step S200 extracts different time-frequency features by performing multi-scale decomposition on the total power of the photovoltaic-storage-charging microgrid to extract the energy features of each layer, combining the physical laws of photovoltaic power to decompose it into trend and periodic components, calculating the power spectral density of photovoltaic power, constructing and normalizing feature vectors containing multi-scale information, and using the time-frequency energy distribution obtained by wavelet transform to calculate time-frequency correlation features, and finally extracting multi-dimensional time-frequency features covering energy, frequency domain and time-frequency dynamic characteristics.

[0084] S300: Based on different time-frequency characteristics, a nonlinear mapping relationship is established through a deep neural network prediction model to determine the optimal network structure and hyperparameter combination;

[0085] In this embodiment of the invention, step S300, which determines the optimal network structure and hyperparameter combination, includes:

[0086] Meteorological, historical power, and calendar features are concatenated into a fusion input vector;

[0087] Based on the fused input vector, a multi-layer nonlinear transformation is defined;

[0088] Based on nonlinear transformation, attention scores are calculated, and attention weights are obtained through normalization.

[0089] Based on the deviation between the attention weights and the true values, the prediction loss is calculated.

[0090] Based on prediction loss, a total loss function is constructed by combining uncertainty loss and regularization loss;

[0091] Based on the total loss function, activation function, optimizer, and learning rate scheduling strategy are set, and the optimal network structure and hyperparameter combination are determined through cross-validation.

[0092] In this embodiment of the invention, the multilayer nonlinear transformation is defined as follows:

[0093] ;

[0094] in, This is the output of the (l+1)th hidden layer. For activation function, Let l be the weight matrix of the l-th layer. For the input of layer l, This is the bias vector.

[0095] In this embodiment of the invention, the fused input vector is represented as:

[0096] ;

[0097] in, To fuse feature vectors, As a meteorological feature, Historical power characteristics, It is a calendar feature.

[0098] In this embodiment of the invention, the attention weight is represented as:

[0099] ;

[0100] in, Let be the attention weight at the s-th time step. The original attention score at time step s is... dummy variables for time steps The attention score, where S is the length of the time series.

[0101] In this embodiment of the invention, the total loss function is expressed as:

[0102] ;

[0103] in, For the total loss, To predict losses, Losses due to uncertainty Here, η represents the regularization loss, and η is the uncertainty loss weight. The weights are for regularization loss.

[0104] In this embodiment of the invention, loss is predicted. Represented as:

[0105] ;

[0106] in, Let be the predicted mean of the k-th sample. Let K be the true power value of the k-th sample, where K is the total number of samples.

[0107] In this embodiment of the invention, the deep neural network prediction model in step S300 includes a high-dimensional feature vector constructed based on multi-scale decomposition, a convolutional neural network (CNN) to extract local spatiotemporal features, a long short-term memory network (LSTM) to capture long-term temporal dependencies, and an attention mechanism to dynamically weight key time steps. The optimal network structure and hyperparameters are determined through cross-validation.

[0108] S400: Based on the optimal network structure and hyperparameter combination, the prediction uncertainty is quantified and the prediction confidence interval is obtained through the Bayesian deep learning framework;

[0109] In this embodiment of the invention, obtaining the predicted confidence interval in step S400 includes:

[0110] By using a Bayesian deep learning framework, parameter uncertainty is expressed as a posterior distribution;

[0111] Based on the posterior distribution, the prediction mean and total prediction variance are calculated to obtain the prediction uncertainty.

[0112] Define uncertainty loss, take the prediction uncertainty as the learning objective, and obtain the prediction confidence interval.

[0113] In this embodiment of the invention, the parameter uncertainty is expressed as a posterior distribution as follows:

[0114] ;

[0115] in, For given data Parameters under conditions The posterior distribution, Let be the likelihood function. For the prior distribution, For marginal likelihood, Let θ be the prior distribution of the parameter θ.

[0116] In this embodiment of the invention, the predicted mean and total predicted variance are calculated as follows:

[0117] ;

[0118] ;

[0119] in, To predict the mean, For network output, This represents the total forecast variance. To observe the noise variance, It is the integral variable.

[0120] In this embodiment of the invention, uncertainty loss is defined. Represented as:

[0121] ;

[0122] in, Let Variance be the prediction variance for the k-th sample. As the average factor, For summation, The square of the prediction error. For the standardized error term, For logarithmic variance regularization term, Let be the predicted mean of the k-th sample. This represents the true power value of the k-th sample.

[0123] In this embodiment of the invention, obtaining the prediction confidence interval in step S400 includes constructing a deep neural network to output the prediction mean and variance, assuming that the prediction result follows a Gaussian distribution to establish a probabilistic output model, and using MC Dropout to perform multiple forward samplings during the inference stage. The mean and variance of the sampling results are used to estimate the expectation and uncertainty of the prediction, and finally the prediction confidence interval is calculated by combining the confidence level.

[0124] In this embodiment of the invention, obtaining the integrated prediction result in step S500 includes:

[0125] Define the weighted average form of the ensemble forecast;

[0126] Based on the weighted average form, the objective function is optimized on the validation set to obtain the optimized weights;

[0127] Based on the optimized weights, the weights are dynamically adjusted according to the real-time root mean square error of each model and the exponential decay law through a dynamic weight update mechanism to obtain the updated weights.

[0128] The integrated prediction results are obtained by normalizing the updated weights.

[0129] In this embodiment of the invention, calculating the total prediction variance includes using Bayesian inference to capture model uncertainty, while introducing an observation noise term to model data uncertainty.

[0130] In this embodiment of the invention, multiple base learners with different structures are trained, including different network architectures and different combinations of input features.

[0131] In this embodiment of the invention, the weighted average form of the integrated prediction is defined as follows:

[0132] ;

[0133] in, For integrated structure expressions, Let m be the total number of base models and m be the index of the base model. Let be the predicted value of the m-th base model at time t. Let be the normalized weights of the m-th model at time t.

[0134] In this embodiment of the invention, the objective function for weight optimization is expressed as:

[0135] ;

[0136] in, This is the actual power value. ζ For regularization parameters, This is an L1 regularization term.

[0137] In this embodiment of the invention, the dynamic weight update mechanism is expressed as follows:

[0138] ;

[0139] in, For the updated unnormalized weights, The unnormalized weights before the update. The attenuation coefficient is... The root mean square error, It is an exponential function. This refers to the dynamic weight update cycle of the ensemble model.

[0140] In this embodiment of the invention, the normalization of the updated weights is represented as follows:

[0141] ;

[0142] in, To normalize the weights, Indexing dummy variables for the model The sum of unnormalized weights.

[0143] S500: Based on the prediction confidence interval, multiple base learners with different structures are trained, and the ensemble prediction result is obtained through the ensemble learning algorithm;

[0144] In this embodiment of the invention, the ensemble learning algorithm in step S500 includes constructing dynamic weights based on the point prediction output and uncertainty estimation of each base learner, weighting and fusing the prediction mean of each model as the ensemble point prediction result, and then generating the final ensemble prediction interval based on the weighted prediction interval, thereby obtaining an ensemble prediction result that balances accuracy and reliability.

[0145] S600: Based on integrated prediction results, a power prediction solution for photovoltaic-storage-charging microgrids is obtained by constructing prediction intervals and combining them with point prediction accuracy indicators.

[0146] In this embodiment of the invention, step S600 includes, during the prediction stage, generating multiple possible prediction paths by sampling model parameters and noise multiple times; and generating prediction intervals by statistically analyzing all sampling results according to quantiles. By combining point prediction accuracy metrics (RMSE, MAPE) and interval evaluation metrics (PICP, PINAW), a power prediction solution for photovoltaic-storage-charging microgrids is obtained.

[0147] In this embodiment of the invention, the generation of multiple possible predicted paths is represented as follows:

[0148] ;

[0149] in, This represents the power prediction for the next h steps in the r-th sampling. For the prediction function, This is the current state. Let r be the model parameters for the r-th sampling. It is random noise.

[0150] In this embodiment of the invention, the generated prediction interval is represented as follows:

[0151] ;

[0152] in, The prediction interval is the confidence level. and These are the lower quantile and the upper quantile, respectively.

[0153] In this embodiment of the invention, the root mean square error (RMSE) is expressed as:

[0154] ;

[0155] In this embodiment of the invention, the mean absolute percentage error (MAPE) is expressed as:

[0156] ;

[0157] In this embodiment of the invention, the predicted interval coverage probability ( ) is represented as:

[0158] ;

[0159] in, This is an indicator function.

[0160] In this embodiment of the invention, the normalized average interval width ( ) is represented as:

[0161] ;

[0162] in, This is the upper bound of the prediction interval for the k-th sample. This is the lower bound of the prediction interval for k samples. This represents the average power.

[0163] In this embodiment of the invention, step S600 further includes implementing an online prediction system to provide real-time power prediction and uncertainty quantification results. An incremental learning mechanism is established to continuously update the model using new observation data. A prediction performance monitoring system is designed to evaluate prediction quality in real time. When prediction performance deteriorates, model retraining is automatically triggered. A visualization system for prediction results is established to provide intuitive decision support for schedulers.

[0164] In embodiments of the present invention, such as Figure 2 As shown, this paper describes the entire process of constructing, training, validating, and applying a power prediction model for a photovoltaic-storage-charging microgrid. Starting with data acquisition and preprocessing, features are extracted through multi-scale wavelet decomposition, a deep neural network is constructed, and Bayesian uncertainty modeling is introduced to quantify prediction uncertainty. Ensemble learning training and prediction model validation are performed. If the prediction accuracy is not up to standard, the network structure parameters are adjusted or the uncertainty model is optimized. If the accuracy is up to standard, further uncertainty quantification assessment and confidence interval rationality judgment are conducted. When the confidence interval is reasonable, the model enters the online prediction deployment stage, and real-time performance monitoring is implemented. During operation, incremental learning updates are performed as needed to maintain the model's accuracy and adaptability, ultimately outputting the prediction results. The entire process embodies a closed-loop mechanism from data-driven approach to model optimization and real-time application, ensuring the high accuracy and robustness of the prediction model.

[0165] Example 2, referring to Tables 1-3, is an embodiment of the present invention. This embodiment provides a multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids. In order to verify the beneficial effects of the present invention, specific implementation methods and implementation effects are scientifically demonstrated.

[0166] The specific details of this embodiment are as follows:

[0167] This study uses a simulation verification of a photovoltaic-storage-charging microgrid system in an industrial park. The system, implemented using MATLAB, comprises a 1.2MW photovoltaic power generation system, an 800kWh energy storage system, and 20 charging piles. The experiment utilizes actual operational data from 2024, including 8760 hours of continuous monitoring data. Meteorological data, sourced from the National Meteorological Administration, includes hourly information on solar radiation, temperature, humidity, and wind speed. Multiple forecast timescales were set, including 15-minute, 1-hour, 4-hour, and 24-hour forecasts.

[0168] Table 1 Comparison of Prediction Performance of Different Methods

[0169]

[0170] As shown in Table 1, this invention significantly outperforms the comparative methods in all evaluation metrics. RMSE is reduced to 72.5kW, a 53.8% improvement compared to the traditional ARIMA method; MAPE is reduced to 5.94%, a 16.6% improvement compared to the best comparative method, Transformer; the predicted interval coverage probability reaches 96.7%, close to the theoretical value of 95%, indicating that the uncertainty quantification is accurate and reliable; the normalized average interval width is only 0.218, demonstrating that narrow interval prediction is achieved while ensuring coverage probability.

[0171] Table 2 Comparison of Prediction Effects at Different Time Scales

[0172]

[0173] Table 2 shows the performance of this invention at different prediction timescales. As the prediction duration increases, the prediction error and uncertainty gradually increase, which is in line with expectations. The 15-minute ultra-short-term prediction has the highest accuracy, with a MAPE of only 3.78%, suitable for real-time control; the 24-hour medium-to-long-term prediction has a MAPE of 15.27%, although the accuracy decreases slightly, it still meets the day-ahead scheduling requirements. The calculation time is controlled within 1 second for all predictions, meeting the requirements of real-time applications.

[0174] Table 3. Analysis of Predictive Effects of Different Power Components

[0175]

[0176] Table 3 analyzes the prediction characteristics and performance of different power components. Energy storage systems exhibit the highest prediction accuracy, thanks to their relatively deterministic control strategy; charging load prediction is the most challenging, primarily due to the high degree of randomness in user behavior; photovoltaic power generation prediction accuracy is moderate, with meteorological factors being the main influencing factor. Through multi-scale decomposition and deep learning modeling, the prediction accuracy of each component reaches a high level.

[0177] Further sensitivity analysis showed that the accuracy of weather forecasts had the most significant impact on photovoltaic power forecasting; a 20% increase in weather forecast error led to an approximately 3.2 percentage point increase in the photovoltaic forecast MAPE. Changes in energy storage control strategies had a relatively small impact on overall forecast accuracy, indicating that the method has good adaptability. Correction analysis of the uncertainty quantification results showed that the actual coverage probability within the forecast interval deviated from the theoretical value by less than 2%, validating the accuracy of the uncertainty estimation.

[0178] The results show that the present invention can significantly improve the power prediction accuracy of photovoltaic-storage-charging microgrids, while providing reliable uncertainty quantification, thus providing important technical support for the safe and economical operation of distribution networks.

[0179] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-scale power prediction and uncertainty quantification method for photovoltaic-storage-charging microgrids, characterized in that, include: Acquire historical operational data and meteorological data for the photovoltaic-storage-charging microgrid; Based on the preprocessed data, different time-frequency features are extracted using a multi-scale decomposition algorithm; Based on different time-frequency characteristics, a nonlinear mapping relationship is established through a deep neural network prediction model to determine the optimal network structure and hyperparameter combination. Based on the optimal network structure and hyperparameter combination, the prediction uncertainty is quantified and the prediction confidence interval is obtained through the Bayesian deep learning framework. Based on the predicted confidence interval, multiple base learners with different structures are trained, and the ensemble prediction result is obtained through an ensemble learning algorithm. Based on the integrated prediction results, a power prediction solution for photovoltaic-storage-charging microgrids is obtained by constructing prediction intervals and combining them with point prediction accuracy indicators. The extraction of different time-frequency features includes: By performing discrete wavelet decomposition on the total power signal of the photovoltaic-storage-charging microgrid, the wavelet detail coefficients and approximation coefficients of each layer are obtained; For photovoltaic power generation, multi-scale components are divided according to physical laws, and decomposed into trend term, daily periodic term and weather disturbance term. By performing wavelet decomposition on the original power signal of the photovoltaic-storage-charging microgrid and calculating the sum of squares of the wavelet detail coefficients of each layer, the energy characteristics of each layer are extracted. Based on the trend term, daily periodic term, and weather disturbance term, the power spectral density of photovoltaic power is extracted by performing a Fourier transform on the photovoltaic power generation. By performing continuous wavelet transform on the total power signal of the system, the time-frequency energy distribution is constructed, and the energy at the same frequency is normalized along the time dimension to obtain the time-frequency correlation characteristics. Based on the energy characteristics, power spectral density, and time-frequency correlation characteristics of each layer, different time-frequency features are extracted. Determining the optimal network structure and hyperparameter combination includes: Meteorological, historical power, and calendar features are concatenated into a fusion input vector; Based on the fused input vector, a multi-layer nonlinear transformation is defined; Based on nonlinear transformation, attention scores are calculated, and attention weights are obtained through normalization. Based on the deviation between the attention weights and the true values, the prediction loss is calculated. Based on prediction loss, a total loss function is constructed by combining uncertainty loss and regularization loss; Based on the total loss function, the activation function, optimizer, and learning rate scheduling strategy are set, and the optimal network structure and hyperparameter combination are determined through cross-validation. The obtained prediction confidence interval includes: By using a Bayesian deep learning framework, parameter uncertainty is expressed as a posterior distribution; Based on the posterior distribution, the predicted mean and the total predicted variance are calculated to obtain the prediction uncertainty. Define uncertainty loss, take the prediction uncertainty as the learning objective, and obtain the prediction confidence interval.

2. The method for multi-scale power prediction and uncertainty quantification of a photovoltaic-storage-charging microgrid as described in claim 1, characterized in that: The obtained integrated prediction results include: Define the weighted average form of the ensemble forecast; Based on the weighted average form, the objective function is optimized on the validation set to obtain the optimized weights; Based on the optimized weights, the weights are dynamically adjusted according to the real-time root mean square error of each model and the exponential decay law through a dynamic weight update mechanism to obtain the updated weights. The integrated prediction results are obtained by normalizing the updated weights.

3. The method for multi-scale power prediction and uncertainty quantification of a photovoltaic-storage-charging microgrid as described in claim 2, characterized in that: The calculation of the total prediction variance includes using Bayesian inference to capture model uncertainty, while introducing an observation noise term to model data uncertainty.

4. The method for multi-scale power prediction and uncertainty quantification of a photovoltaic-storage-charging microgrid as described in claim 3, characterized in that: The uncertainty loss is defined to include the variance weighting term of the prediction error and the logarithmic regularization term of the prediction variance.

5. The method for multi-scale power prediction and uncertainty quantification of a photovoltaic-storage-charging microgrid as described in claim 4, characterized in that: The dynamic weight update mechanism is expressed as follows: in, For the updated unnormalized weights, The unnormalized weights before the update. The attenuation coefficient is... The root mean square error, It is an exponential function. This refers to the dynamic weight update cycle of the ensemble model.

6. The method for multi-scale power prediction and uncertainty quantification of a photovoltaic-storage-charging microgrid as described in claim 5, characterized in that: The time-frequency correlation feature is represented as follows: in, It is a time-frequency correlation. To charge the total power of the photovoltaic-storage microgrid, The translation parameter in continuous wavelet transform. For wavelet transform, For integration time variable, Integrating along the time axis, The variable is time, representing the moment when the power signal changes over time. For frequency variables, it represents the frequency components of the signal in the frequency domain.

7. The method for multi-scale power prediction and uncertainty quantification of a photovoltaic-storage-charging microgrid as described in claim 6, characterized in that: The uncertainty loss Represented as: in, Let K be the prediction variance of the k-th sample, and K be the total number of samples. As the average factor, For summation, The square of the prediction error. For the standardized error term, For logarithmic variance regularization term, Let the predicted mean of the k-th sample be . This represents the true power value of the k-th sample.

Images