Photovoltaic mechanism electricity price dynamic prediction method and system based on neural network model

Abstract

The invention provides a photovoltaic mechanism electricity price dynamic prediction method and system based on a neural network model, and belongs to the technical field of electricity market electricity price prediction, and the method comprises the following steps: collecting multi-voltage level photovoltaic power generation data, user electricity consumption data and multi-source external feature data of a target regional power grid, and forming a time sequence data set; preprocessing the time sequence data set to obtain a core feature set; respectively constructing an LSTM basic prediction model and a VMD-PLO-Transform-LSTM mixed prediction model, and training by adopting the core feature set to obtain a dual-prediction model; and calculating an initial predicted value of the dual-prediction model, and weighting and fusing the initial predicted value by taking the real-time mean square error reciprocal of the two models as a weight to obtain a final photovoltaic mechanism electricity price predicted value. According to the photovoltaic mechanism electricity price dynamic prediction method based on the neural network model provided by the invention, a high-precision dynamic prediction result can be output based on double-model cooperation of the LSTM basic model and the VMD-PLO-Transform-LSTM mixed model.

Description

Technical Field

[0001] This invention belongs to the field of electricity market price prediction technology, and more specifically, it relates to a method and system for dynamic prediction of photovoltaic power prices based on a neural network model. Background Technology

[0002] With the acceleration of the global energy transition, my country's installed capacity of new energy (especially photovoltaic power generation) has continued to grow. In 2023, the national photovoltaic power generation exceeded 500 billion kilowatt-hours, accounting for more than 6% of the country's total power generation. To promote the consumption of new energy and the market-oriented reform of the electricity market, Hebei Province took the lead in introducing a policy to reform the on-grid tariff for new energy, clarifying that from December 2025, a dual-track system of "market formation + mechanism-based tariff guarantee" will be fully implemented. That is, the on-grid tariff for photovoltaic power will be formed primarily through market transactions. When the market tariff is lower than the mechanism-based tariff guarantee, subsidies will be provided to new energy enterprises through a "refund for overpayment and supplement for underpayment" settlement method. When the market tariff is higher than the mechanism-based tariff guarantee, the excess will be returned to electricity users proportionally, thereby balancing market fluctuations and the interests of both enterprises and users.

[0003] Against this backdrop, the accuracy of dynamic forecasting of mechanism-based electricity prices directly determines the fairness and effectiveness of price difference settlement: if the forecast is too high, it will lead to excessive subsidies from grid companies; if the forecast is too low, it will harm the revenue of new energy companies; if the forecast is lagging, it will be impossible to adjust dispatch strategies in a timely manner, causing an imbalance between power supply and demand. Therefore, constructing a high-precision photovoltaic mechanism-based electricity price forecasting method adapted to the market characteristics of Hebei Province has become a core requirement for current power market operation. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for dynamic prediction of photovoltaic power prices based on a neural network model, in order to meet market demands.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: to provide a method for dynamic prediction of photovoltaic power generation prices based on a neural network model, comprising the following steps: S1. Collect photovoltaic power generation data, user electricity consumption data and multi-source external feature data of the target area power grid at multiple voltage levels to form a time-series dataset; S2. Perform data preprocessing on the time-series dataset to obtain the core feature set; S3. Construct the LSTM basic prediction model and the VMD-PLO-Transformer-LSTM hybrid prediction model respectively, and train the two models using the core feature set to obtain the trained dual prediction model. S4. Calculate the preliminary prediction value of the dual prediction model, and then use the reciprocal of the real-time mean square error of the dual prediction model as the weight to obtain the final photovoltaic power price prediction value by weighted fusion of the preliminary prediction value of the dual prediction model.

[0006] In one possible implementation, the multi-source external feature data in step S1 includes: meteorological feature data, power grid operation feature data, and policy regulation feature data.

[0007] In one possible implementation, the data preprocessing in step S2 specifically includes: S21. Outlier removal: The 3σ criterion is used to process data that conforms to a normal distribution, and the interquartile range method is used to process data that is skewed. The outlier determination threshold of the interquartile range method is [Q1-1.5×IQR,Q3+1.5×IQR], where Q1 is the first quartile and Q3 is the third quartile. S22. Data missing value imputation: For continuous data, linear interpolation is used to imput short-term missing data, where the missing duration of the short-term missing data is ≤4 hours; KNN interpolation is used to imput long-term missing data, where the missing duration of the long-term missing data is >4 hours. S23. Data standardization: The Z-score standardization formula is used to eliminate dimensions from the data. The calculation formula is as follows: z , where x is the original data, μ is the mean of the data class, and σ is the standard deviation of the data class; S24. Feature data screening: An improved mutual information feature selection method combined with Pearson correlation coefficient is adopted to retain features with mutual information value ≥0.6 and Pearson correlation coefficient ≥0.75, forming the core feature dataset.

[0008] In one possible implementation, the LSTM base prediction model in step S3 includes: Input layer, input layer dimension = number of core feature data × time window length, the time window length is set to 24 hours; The hidden layer has 3 layers, with 128 nodes in each layer. The first hidden layer uses the ReLU activation function, the second layer uses the tanh activation function, and the third layer uses the ReLU activation function. The output layer uses a linear activation function to output the predicted electricity price for a single time period. A Dropout layer with a dropout rate of 0.2 is set between the first and second hidden layers; an L2 regularization term with a regularization coefficient of 0.001 is set before the output layer. The training of the LSTM basic prediction model includes: using the Adam optimizer, setting the initial learning rate to 0.001, and decaying the learning rate to 0.8 times the current value after every 20 training rounds; using mean squared error as the loss function; using a training batch size of 32, a training cycle of 100 rounds, and adopting an early stopping strategy, stopping training when the validation set loss does not decrease for 10 consecutive rounds.

[0009] In one possible implementation, the construction and training of the VMD-PLO-Transformer-LSTM hybrid prediction model in step S3 includes: S321. Electricity price series decomposition: The historical electricity price series is decomposed into multiple stationary intrinsic mode functions using variational mode decomposition. S322, Transformer feature extraction unit setting, the input layer is a decomposed single stationary intrinsic mode function sequence, the encoder is set with 4 layers, each layer contains a multi-head attention mechanism and a feedforward neural network, the number of attention heads is optimized by the Aurora optimization algorithm; S323, LSTM local modeling unit settings, input is the feature vector output by Transformer, set 3 LSTM layers, each layer has 64 nodes, and the forget gate threshold is optimized by Aurora optimization algorithm; S324. Model training: The objective function of the Aurora optimization algorithm is set as follows: Where α=0.8 is the error weight, MSE is the mean squared error, β=0.2 is the computation time weight, and T is the time taken for a single prediction by the model; the Aurora optimization algorithm iterates 50 times, and after optimization, it is trained with the same training parameters as the LSTM basic prediction model to obtain the prediction sub-models of each stationary intrinsic mode function.

[0010] In one possible implementation, the weighted fusion described in step S4 specifically includes: S41. Real-time error calculation: Collect actual electricity price data hourly and calculate the real-time mean square error of the LSTM basic prediction model. Real-time mean square error of the VMD-PLO-Transformer-LSTM hybrid prediction model ; S42. Weight update: Calculate dynamic weights based on real-time mean square error. , ,in Weights for the LSTM base prediction model. Weights for the VMD-PLO-Transformer-LSTM hybrid prediction model; S43. The final formula for the predicted photovoltaic power price, obtained through fusion calculation, is: ,in These are the initial predictions output by the LSTM base prediction model. The initial predictions output by the VMD-PLO-Transformer-LSTM hybrid prediction model.

[0011] In one possible implementation, the meteorological characteristic data includes daily sunshine intensity, ambient temperature, wind speed, and precipitation; the power grid operation characteristic data includes regional total load factor, transmission line utilization rate, and energy storage device charging and discharging status; and the policy regulation characteristic data includes the mechanism-based electricity price adjustment coefficient and the timeliness factor of new energy subsidy policies.

[0012] In one possible implementation, the extraction of the policy regulation feature data includes: S111. Extraction of the adjustment coefficient for the safety net electricity price: Obtain historical records of the safety net electricity price adjustment from the electricity price policy documents issued by the Development and Reform Commission of the target region, and use linear interpolation to complete the adjustment coefficient k for non-adjustment days. The adjustment coefficient k satisfies the following conditions: when the safety net electricity price is increased by 5%, k=1.05; when the safety net electricity price remains unchanged, k=1.00. S112. Extraction of the timeliness factor of the new energy subsidy policy: Based on the effective date and the expiration date of the subsidy policy, construct the timeliness factor t: before the policy takes effect t=0, during the effective period t=1, after the expiration t=0.8, and after the complete withdrawal t=0. S113. Policy feature standardization: Min-Max standardization is used to map k and t to the [0,1] interval, and unify the dimensions with other external features.

[0013] 9. In one possible implementation, the specific steps for improving the mutual information feature selection in step S24 include: S241. Calculate the initial mutual information. Based on historical data, calculate the mutual information value between each candidate feature and the electricity price sequence. Where X is a candidate feature and Y is the electricity price sequence; S242. Redundancy correction: Calculate the mutual information value between any two candidate features. Correct the initial mutual information: where n is the total number of candidate features; S243, Feature sorting and filtering, according to... Sort in descending order and retain The core feature set is obtained by taking the intersection of features with a correlation coefficient ≥0.6 and features with a Pearson correlation coefficient ≥0.75.

[0014] The beneficial effects of the photovoltaic power price dynamic prediction method based on a neural network model provided by this invention are as follows: Compared with existing technologies, this invention provides a dynamic prediction method for photovoltaic (PV) electricity prices based on a neural network model. By collecting multi-source data including PV data at multiple voltage levels, electricity consumption data of different user types, and data on meteorology, grid operation, and policy regulation, it overcomes the shortcomings of traditional methods in terms of incomplete data coverage. Through differentiated outlier removal, targeted missing value imputation, and improved feature selection combining mutual information and Pearson correlation coefficient, it enhances data quality to ensure the reliability of model input. Relying on the synergy of a dual-model approach—a basic LSTM model and a VMD-PLO-Transformer-LSTM hybrid model—it solves the problem of poor adaptability of a single model and balances prediction accuracy and time consumption. Furthermore, it uses real-time reciprocal mean square error to dynamically weight and fuse the results of the two models, avoiding the accumulation of errors with fixed weights, while simultaneously achieving quantitative modeling of policy regulation factors. Finally, it outputs a high-precision, dynamic prediction value for PV electricity prices.

[0015] The present invention also provides a photovoltaic power generation price dynamic prediction system based on a neural network model, which adopts the above-mentioned photovoltaic power generation price dynamic prediction method based on a neural network model.

[0016] The beneficial effects of the photovoltaic power price dynamic prediction system based on a neural network model provided by this invention are as follows: Compared with existing technologies, the photovoltaic mechanism electricity price dynamic prediction system based on the neural network model of this invention adopts the above-mentioned prediction method, which can integrate multi-source data including photovoltaic data of multiple voltage levels, user electricity consumption, meteorological data, power grid operation data, and policy regulation data. The data quality is guaranteed by preprocessing such as differentiated outlier removal, targeted missing value filling, and core feature screening. Relying on dual-model collaborative training and dynamic weighted fusion, it avoids the defects of incomplete data, poor model adaptation, fixed weights, and difficulty in quantifying policies in traditional prediction. It can output high-precision dynamic prediction results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0018] Figure 1 A flowchart of a photovoltaic mechanism electricity price dynamic prediction method based on a neural network model provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0020] Please see Figure 1 The present invention will now describe the method for dynamic prediction of photovoltaic power generation prices based on a neural network model. The method includes the following steps: S1. Collect photovoltaic power generation data, user electricity consumption data and multi-source external feature data of the target area power grid at multiple voltage levels to form a time-series dataset; S2. Perform data preprocessing on the time-series dataset to obtain the core feature set; S3. Construct the LSTM basic prediction model and the VMD-PLO-Transformer-LSTM hybrid prediction model respectively, and train the two models using the core feature set to obtain the trained dual prediction model. S4. Calculate the preliminary prediction value of the dual prediction model, and then use the reciprocal of the real-time mean square error of the dual prediction model as the weight to obtain the final photovoltaic power price prediction value by weighted fusion of the preliminary prediction value of the dual prediction model.

[0021] In the application, during step S1, hourly data of the target area's power grid for at least the past three years should be collected. This includes multi-voltage level photovoltaic power generation data such as hourly output power, power generation, and curtailment rate of photovoltaic power plants at different voltage levels (110kV, 220kV, 500kV, etc.) within the target area. Different voltage levels correspond to photovoltaic power plants of different scales, and their output fluctuations have varying degrees of impact on electricity prices. Multi-voltage data can more comprehensively reflect the photovoltaic supply capacity. User electricity consumption data is hourly load data categorized by user type (industrial, residential, commercial). Electricity demand is the core driver of electricity prices; categorized data can improve the accuracy of demand-side characteristics.

[0022] In step S1, the multi-source external characteristic data includes meteorological characteristic data, power grid operation characteristic data, and policy regulation characteristic data. Among them, meteorological characteristic data directly affects the output of photovoltaic power plants (e.g., solar irradiance determines the upper limit of photovoltaic output, and temperature affects the efficiency of photovoltaic modules). In this step, the meteorological characteristic data includes daily solar irradiance (unit: W / m²), ambient temperature (unit: °C), wind speed (unit: m / s), and precipitation (unit: mm), which are all core meteorological factors affecting photovoltaic output.

[0023] Power grid operation characteristic data reflects the power grid's supply and demand balance (e.g., excessively high total load factor may lead to higher electricity prices, while excessively low transmission line utilization may result in wasted electricity). Power grid operation characteristic data includes regional total load factor (actual load / maximum load), transmission line utilization rate (actual transmission power / rated transmission power), and the charging and discharging status of energy storage devices (charging is negative, discharging is positive), reflecting the power grid's supply and demand and operational efficiency. A higher total load factor indicates stronger electricity demand, potentially leading to higher electricity prices; excessively low transmission line utilization results in wasted grid capacity, while excessively high utilization may trigger overload protection; energy storage increases electricity demand during charging and supplements power supply during discharging, directly affecting electricity price fluctuations.

[0024] Policy regulation characteristic data are core regulatory factors under the dual-track system (e.g., adjustments to the safety net electricity price directly determine the subsidy benchmark, and the timeliness of subsidy policies affects the power generation willingness of new energy enterprises). Policy regulation characteristic data includes the adjustment coefficient of the safety net electricity price mechanism and the timeliness factor of new energy subsidy policies, which are directly related to the electricity price forecast benchmark.

[0025] In this step, the method for extracting the policy regulation feature data includes: S111. Extraction of the mechanism-based safety net electricity price adjustment coefficient: Historical safety net electricity price adjustment records are obtained from the electricity price policy documents issued by the Development and Reform Commission of the target region (such as the "Notice on Adjustment of New Energy On-grid Electricity Price in Hebei Province in 2024"). When the safety net electricity price increases by 5%, k=1.05; when it increases by 10%, k=1.10; when the safety net electricity price remains unchanged, k=1.00; when it decreases by 5%, k=0.95. On non-adjustment days (dates without policy document issuance), the k value is supplemented using linear interpolation. For example, if k=1.00 on January 1, 2024, and k=1.05 on March 1, 2024, then the k value from January 1 to March 1 gradually increases from 1.00 to 1.05 according to a linear trend, ensuring the temporal continuity of policy characteristics.

[0026] S112. Extraction of the timeliness factor for new energy subsidy policies: Based on the effective date and expiration date of the subsidy policy, construct the timeliness factor t: t=0 before the policy takes effect, t=1 during the policy's effective period, t=0.8 after the policy expires (transition period), and t=0 after the policy is completely withdrawn. For example, if a subsidy policy takes effect on January 1, 2024, expires on December 31, 2024, enters the transition period on January 1, 2025, and is completely withdrawn on June 1, 2025, then t=0 before December 31, 2023, t=1 from January to December 2024, t=0.8 from January to May 2025, and t=0 after June 2025.

[0027] S113. Policy feature standardization: Min-Max standardization is used to map k and t to the [0,1] interval, as shown in the formula. For example, if the historical maximum value of k is 1.10 and the minimum value is 0.95, then k´=(k-0.95) / (1.10-0.95), ensuring that the dimensions of policy characteristics are consistent with those of other external characteristics.

[0028] In application, the data preprocessing of the time series dataset in step S2 specifically includes: S21. Data outlier removal: Differentiated methods are used for data with different distributions. For data that conforms to a normal distribution (such as photovoltaic power output data under normal weather conditions and residential electricity load data), the 3σ criterion is used.

[0029] When data values ​​exceed [μ-3σ, μ+3σ] (where μ is the mean and σ is the standard deviation), they are identified as outliers and removed. For skewed data (such as industrial electricity consumption data under extreme weather conditions or photovoltaic power output data after heavy rain), the interquartile range (IQR) method is used. The outlier threshold for the IQR method is [Q1-1.5×IQR, Q3+1.5×IQR], where Q1 is the first quartile, Q3 is the third quartile, and IQR = Q3-Q1. This step effectively removes outliers caused by equipment failures and data acquisition errors, preventing them from interfering with model training.

[0030] S22. Data missing value imputation: Classified by the duration of the missing data. For short-term missing data (missing time ≤ 4 hours), linear interpolation is used (e.g., 2 hours of missing data due to a brief communication interruption). The data is imputed by the linear trend of the data before and after the missing segment to ensure the continuity of the data sequence. For long-term missing data (missing time > 4 hours), KNN interpolation is used (e.g., 12 hours of missing data due to a photovoltaic power station equipment failure). The feature similarity between the missing sample and other samples is calculated, and the mean of the k most similar samples (k is usually 5-10) is selected to impute the missing data, thereby improving the imputation accuracy.

[0031] S23. Data standardization: The Z-score standardization formula is used to eliminate dimensions from the data. The calculation formula is as follows: z , where x is the original data, μ is the mean of the data, and σ is the standard deviation of the data. For example, light intensity (unit: W / m², numerical range 0-1000) and temperature (unit: ℃, numerical range -10-40) have significantly different dimensions. Standardization can map all features to similar numerical ranges, avoiding the model from being overly biased towards features with large numerical values.

[0032] S24. Feature data screening: An improved mutual information feature selection method combined with Pearson correlation coefficient is adopted to retain features with mutual information value ≥0.6 and Pearson correlation coefficient ≥0.75, forming the core feature dataset.

[0033] In step S24, the specific steps for improving the mutual information feature selection include: S241. Calculate the initial mutual information. Based on historical data, calculate the mutual information value between each candidate feature X (such as light intensity, total load factor) and the electricity price sequence Y. Where X is the candidate feature and Y is the electricity price sequence. The larger the value, the stronger the correlation between feature X and electricity price Y.

[0034] S242. Redundancy correction: Calculate any two candidate features. and Mutual information values ​​between The initial mutual information is corrected to obtain the corrected mutual information value. ,in, where n is the total number of candidate features; S243, Feature sorting and filtering, according to... Sort in descending order and retain The core feature set is obtained by taking the intersection of features with a correlation coefficient ≥0.6 and features with a Pearson correlation coefficient ≥0.75.

[0035] In step S3, the structural parameters of the LSTM base prediction model are as follows: Input layer: Dimension = Number of core feature data × Time window length (The time window is set to 24 hours because electricity prices have obvious intraday cyclical patterns, and a 24-hour window can fully capture intraday fluctuation features). Hidden layers: 3 layers, 128 nodes per layer (balancing fitting ability and computational cost). The activation function combination is "ReLU-tanh-ReLU": the first and third layers use ReLU activation function to solve the gradient vanishing problem and improve the model's ability to fit non-linear features; the second layer uses tanh activation function to map the output to [-1,1], enhancing the model's sensitivity to data fluctuations. Regularization is applied to the hidden layers. A Dropout layer (dropout rate = 0.2) is placed between the first and second hidden layers to randomly drop 20% of neurons, preventing overfitting. An L2 regularization term (regularization coefficient = 0.001) is placed before the output layer: by penalizing the absolute value of the model weights, overfitting is further suppressed. Output layer: A linear activation function is used to output the predicted electricity price for a single time period (1 hour) (since the electricity price is a continuous value, the linear activation function is suitable for regression tasks).

[0036] The training method for the LSTM basic prediction model includes: using the Adam optimizer, setting the initial learning rate to 0.001, and decaying the learning rate to 0.8 times the current value after every 20 training rounds; using mean squared error as the loss function; using a training batch size of 32, a training cycle of 100 rounds, and adopting an early stopping strategy, stopping training when the validation set loss does not decrease for 10 consecutive rounds.

[0037] In step S3, the construction and training of the VMD-PLO-Transformer-LSTM hybrid prediction model includes: S321. Electricity price series decomposition: Variational mode decomposition (VMD) is used to decompose historical electricity price series into multiple stationary intrinsic mode functions (IMFs). Electricity price series usually contain components such as "intraday high-frequency fluctuations (such as short-term load changes), intraweekly medium-frequency fluctuations (such as differences between weekdays and weekends), and monthly low-frequency fluctuations (such as seasonal changes)". Directly modeling non-stationary series can easily lead to error accumulation. VMD can decompose non-stationary series into multiple IMFs with fixed frequencies and stable fluctuations through iterative optimization. Modeling each IMF separately can improve prediction accuracy. The S322 Transformer feature extraction unit is configured with a single decomposed IMF sequence as input. For long-term dependency features of the IMF sequence (such as cross-day electricity price correlations), a four-layer encoder is used. Each layer includes a multi-head attention mechanism and a feedforward neural network. Multiple attention heads are used to compute the correlation weights between features in parallel, capturing long-term dependencies in different dimensions (such as the correlation between a specific IMF and IMFs from the same period three days prior). The feedforward neural network performs a non-linear transformation on the features output by the attention mechanism, enhancing feature representation capabilities. The number of attention heads is optimized using the Aurora Optimization (PLO) algorithm. PLO is an intelligent optimization algorithm based on the motion of aurora particles in nature, with a faster convergence speed than Particle Swarm Optimization (PSO), and can find the optimal number of attention heads (typically 4-8 heads), avoiding the subjectivity of manual setting.

[0038] The S323 LSTM local modeling unit setting supplements the short-term dependency modeling capability for the features extracted by the Transformer. The input is the feature vector output by the Transformer. It is set with 3 LSTM layers, each with 64 nodes. The forget gate threshold is optimized by PLO. If the forget gate threshold is too large, historical information will be lost; if it is too small, information will be redundant. PLO can find a balance threshold (usually 0.3-0.5).

[0039] S324. Model training: The objective function of the Aurora optimization algorithm is set as follows: Where α=0.8 is the error weight, MSE is the mean squared error, β=0.2 is the computation time weight, and T is the time taken for a single prediction by the model; the Aurora optimization algorithm iterates 50 times, and after optimization, it is trained with the same training parameters as the LSTM basic prediction model to obtain the prediction sub-models of each stationary intrinsic mode function.

[0040] In application, the weighted fusion described in step S4 specifically includes: S41. Real-time error calculation: Collect actual electricity price data hourly and calculate the real-time mean square error of the LSTM basic prediction model. Real-time mean square error of the VMD-PLO-Transformer-LSTM hybrid prediction model ; S42. Weight update: Calculate dynamic weights based on real-time mean square error. , ,in Weights for the LSTM base prediction model. Weights for the VMD-PLO-Transformer-LSTM hybrid prediction model; S43. The final formula for the predicted photovoltaic power price, obtained through fusion calculation, is: ,in These are the initial predictions output by the LSTM base prediction model. The initial predictions output by the VMD-PLO-Transformer-LSTM hybrid prediction model.

[0041] The present invention provides a method for dynamic prediction of photovoltaic (PV) electricity prices based on a neural network model. Compared with existing technologies, this method compensates for the incomplete data coverage of traditional methods by collecting multi-source data, including PV data at multiple voltage levels, electricity consumption data of different user types, and data on meteorology, grid operation, and policy regulation. It improves data quality to ensure the reliability of model input through differentiated outlier removal, targeted missing value imputation, and improved feature selection combining mutual information and Pearson correlation coefficient. It relies on the synergy of a dual-model approach, combining a basic LSTM model and a VMD-PLO-Transformer-LSTM hybrid model, to address the poor adaptability of a single model and balance prediction accuracy and time consumption. Furthermore, it uses real-time reciprocal mean square error to dynamically weight and fuse the results of the two models, avoiding the accumulation of errors with fixed weights and simultaneously achieving quantitative modeling of policy regulation factors. Finally, it outputs a high-precision, dynamic PV electricity price prediction.

[0042] This invention also provides a dynamic prediction system for photovoltaic power prices based on a neural network model, which employs the above-mentioned prediction method and includes: Data acquisition module: includes data acquisition terminal (deployed in photovoltaic power plants, user-side electricity meters, and weather stations), communication module (5G / fiber optic), and software including data interface (connecting to the National Development and Reform Commission policy database and the power grid dispatch center database), to realize real-time acquisition and storage of photovoltaic data at multiple voltage levels, user electricity consumption data, and multi-source external characteristic data; Data preprocessing module: Deployed on a cloud server, it is used to preprocess time series datasets and output core feature sets; Model building and training module: Deployed on a GPU server, used for building, training and optimizing parameters of the LSTM base model and the VMD-PLO-Transformer-LSTM hybrid prediction model; Weighted fusion module: Used to call the preliminary prediction values ​​of the LSTM basic model and the VMD-PLO-Transformer-LSTM hybrid prediction model in real time, and perform weighted fusion calculations to output the final prediction value.

[0043] The photovoltaic power price dynamic prediction system based on a neural network model provided by this invention, compared with the prior art, adopts the above-mentioned prediction method, which can integrate multi-source data including photovoltaic data of multiple voltage levels, user electricity consumption, meteorological data, power grid operation data, and policy regulation data. The data quality is ensured through preprocessing such as differentiated outlier removal, targeted missing value filling, and core feature screening. Relying on dual-model collaborative training and dynamic weighted fusion, it avoids the defects of incomplete data, poor model adaptation, fixed weights, and difficulty in quantifying policies in traditional prediction. It can output high-precision dynamic prediction results.

[0044] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic prediction of photovoltaic power prices based on a neural network model, characterized in that, Includes the following steps: S1. Collect photovoltaic power generation data, user electricity consumption data and multi-source external feature data of the target area power grid at multiple voltage levels to form a time-series dataset; S2. Perform data preprocessing on the time-series dataset to obtain the core feature set; S3. Construct an LSTM basic prediction model and a VMD-PLO-Transformer-LSTM hybrid prediction model respectively, and train the two models using the core feature set to obtain a trained dual prediction model; S4. Calculate the preliminary prediction value of the dual prediction model, and then use the inverse of the real-time mean square error of the dual prediction model as the weight to obtain the final photovoltaic power generation price prediction value by weighted fusion of the preliminary prediction value of the dual prediction model.

2. The photovoltaic power price dynamic prediction method based on a neural network model as described in claim 1, characterized in that, The multi-source external feature data mentioned in step S1 includes: meteorological feature data, power grid operation feature data, and policy regulation feature data.

3. The photovoltaic power price dynamic prediction method based on a neural network model as described in claim 1, characterized in that, The data preprocessing described in step S2 specifically includes: S21. Outlier removal: The 3σ criterion is used to process data that conforms to a normal distribution, and the interquartile range method is used to process data that is skewed. The outlier determination threshold of the interquartile range method is [Q1-1.5×IQR,Q3+1.5×IQR], where Q1 is the first quartile and Q3 is the third quartile. S22. Data missing value imputation: For continuous data, linear interpolation is used to imput short-term missing data, where the missing duration of the short-term missing data is ≤4 hours; KNN interpolation is used to imput long-term missing data, where the missing duration of the long-term missing data is >4 hours. S23. Data standardization: The Z-score standardization formula is used to eliminate dimensions from the data. The calculation formula is as follows: z , where x is the original data, μ is the mean of the data class, and σ is the standard deviation of the data class; S24. Feature data screening: An improved mutual information feature selection method combined with Pearson correlation coefficient is adopted to retain features with mutual information value ≥0.6 and Pearson correlation coefficient ≥0.75, forming the core feature dataset.

4. The photovoltaic power price dynamic prediction method based on a neural network model as described in claim 1, characterized in that, The LSTM basic prediction model mentioned in step S3 includes: Input layer, input layer dimension = number of core feature data × time window length, the time window length is set to 24 hours; The hidden layer has 3 layers, with 128 nodes in each layer. The first hidden layer uses the ReLU activation function, the second layer uses the tanh activation function, and the third layer uses the ReLU activation function. The output layer uses a linear activation function to output the predicted electricity price for a single time period. A Dropout layer with a dropout rate of 0.2 is set between the first and second hidden layers; an L2 regularization term with a regularization coefficient of 0.001 is set before the output layer. The training of the LSTM basic prediction model includes: using the Adam optimizer, setting the initial learning rate to 0.001, and decaying the learning rate to 0.8 times the current value after every 20 training rounds; using mean squared error as the loss function; using a training batch size of 32, a training cycle of 100 rounds, and adopting an early stopping strategy, stopping training when the validation set loss does not decrease for 10 consecutive rounds.

5. The photovoltaic power price dynamic prediction method based on a neural network model as described in claim 4, characterized in that, The construction and training of the VMD-PLO-Transformer-LSTM hybrid prediction model in step S3 includes: S321. Electricity price series decomposition: The historical electricity price series is decomposed into multiple stationary intrinsic mode functions using variational mode decomposition. S322, Transformer feature extraction unit setting, the input layer is a decomposed single stationary intrinsic mode function sequence, the encoder is set with 4 layers, each layer contains a multi-head attention mechanism and a feedforward neural network, the number of attention heads is optimized by the Aurora optimization algorithm; S323, LSTM local modeling unit settings, input is the feature vector output by Transformer, set 3 LSTM layers, each layer has 64 nodes, and the forget gate threshold is optimized by Aurora optimization algorithm; S324. Model training: The objective function of the Aurora optimization algorithm is set as follows: Where α=0.8 is the error weight, MSE is the mean squared error, β=0.2 is the computation time weight, and T is the time taken for a single prediction by the model; the Aurora optimization algorithm iterates 50 times, and after optimization, it is trained with the same training parameters as the LSTM basic prediction model to obtain the prediction sub-models of each stationary intrinsic mode function.

6. The photovoltaic power price dynamic prediction method based on a neural network model as described in claim 1, characterized in that, The weighted fusion described in step S4 specifically includes: S41. Real-time error calculation: Collect actual electricity price data hourly and calculate the real-time mean square error of the LSTM basic prediction model. Real-time mean square error of the VMD-PLO-Transformer-LSTM hybrid prediction model ; S42. Weight update: Calculate dynamic weights based on real-time mean square error. , ,in Weights for the LSTM base prediction model. Weights for the VMD-PLO-Transformer-LSTM hybrid prediction model; S43. The final formula for the predicted photovoltaic power price, obtained through fusion calculation, is: ,in These are the initial predictions output by the LSTM base prediction model. The initial predictions output by the VMD-PLO-Transformer-LSTM hybrid prediction model.

7. The photovoltaic power price dynamic prediction method based on a neural network model as described in claim 2, characterized in that, The meteorological characteristic data includes daily sunshine intensity, ambient temperature, wind speed, and precipitation; the power grid operation characteristic data includes regional total load factor, transmission line utilization rate, and energy storage device charging and discharging status; the policy regulation characteristic data includes the mechanism-based electricity price adjustment coefficient and the timeliness factor of new energy subsidy policies.

8. The photovoltaic power price dynamic prediction method based on a neural network model as described in claim 7, characterized in that, The extraction of policy regulation feature data includes: S111. Extraction of the adjustment coefficient for the safety net electricity price: Obtain historical records of the safety net electricity price adjustment from the electricity price policy documents issued by the Development and Reform Commission of the target region, and use linear interpolation to complete the adjustment coefficient k for non-adjustment days. The adjustment coefficient k satisfies the following conditions: when the safety net electricity price is increased by 5%, k=1.05; when the safety net electricity price remains unchanged, k=1.

00. S112. Extraction of the timeliness factor of the new energy subsidy policy: Based on the effective date and the expiration date of the subsidy policy, construct the timeliness factor t: before the policy takes effect t=0, during the effective period t=1, after the expiration t=0.8, and after the complete withdrawal t=0. S113. Policy feature standardization: Min-Max standardization is used to map k and t to the [0,1] interval, and unify the dimensions with other external features.

9. The method for dynamic prediction of photovoltaic power prices based on a neural network model as described in claim 3, characterized in that, The specific steps for improving mutual information feature selection in step S24 include: S241. Calculate the initial mutual information. Based on historical data, calculate the mutual information value between each candidate feature and the electricity price sequence. Where X is a candidate feature and Y is the electricity price sequence; S242. Redundancy correction: Calculate the mutual information value between any two candidate features. Correct the initial mutual information: where n is the total number of candidate features; S243, Feature sorting and filtering, according to... Sort in descending order and retain The core feature set is obtained by taking the intersection of features with a correlation coefficient ≥0.6 and features with a Pearson correlation coefficient ≥0.

75.

10. A photovoltaic power price dynamic prediction system based on a neural network model, characterized in that, The method for dynamic prediction of photovoltaic power price based on a neural network model, as described in any one of claims 1-9, was adopted.

Images