AI-based power curve optimization method of new energy field station for spot market
Through AI-based new energy station power curve optimization method for spot market, multi-factor prediction and multiple models, the subjectivity and accuracy of power curve adjustment in the existing technology are solved, and more robust and efficient electricity trading decisions.
Patent Information
- Application Number
- CN202410943486.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-15
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing power curve adjustment methods have strong subjectivity, time-consuming, accurate and consistent challenges, and fail to effectively consider the uncertainty of the prediction model, resulting in poor efforts in the complex and rapidly changing power market.
Using AI-based new energy station power curve for spot market, we can use the method of finding the best power curve for the spot market to predict the price difference between the clearing price and the real-time clearing price, combined with the comprehensive consideration of multiple factors, establish an optimal declaration curve model, emphasizing intelligent adjustment and risks Control, use machine learning and multiple models (such as XGBoost, LightGBM, LSTM, K-means, Stacking models) for data processing and prediction, avoid randomness, and optimize electricity and energy bills.
It improves the accuracy and stability of power curve adjustment, enhances the adaptability and reliability of the model, reduces the risks and labor costs of power transaction decisions, and provides more advanced and reliable solutions.
Smart Images

Figure CN119005386A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power market, and in particular to a power curve optimization method for a new energy station based on AI for a spot market. Background Art
[0002] At present, the power trading market, as a key component of the energy industry, faces a series of challenges and shortcomings. The main problem is that the existing power curve adjustment method has certain limitations. At present, the adjustment of the power curve is usually performed manually by transaction-related personnel, which is not only highly subjective, time-consuming, labor-intensive, and costly, but also has challenges in the accuracy and consistency of the adjustment. These shortcomings make the current power curve adjustment method seem incapable of coping with the complex and rapidly changing power market.
[0003] In the prior art, the prediction of price difference is mostly focused on the difference between the real-time electricity price and the day-ahead electricity price, and the price difference is obtained by simple subtraction. However, this method has a significant defect, that is, it introduces two sources of uncertainty. At the same time, the prediction errors of the real-time electricity price and the day-ahead electricity price cause the final price difference result to show a certain degree of randomness. The deficiency of the prior art in dealing with this randomness is that they do not fully consider the uncertainty brought by each prediction model, and thus fail to effectively control risks.
[0004] To solve the above problems, the present invention introduces more comprehensive prediction factors, including site power prediction, day-ahead spot electricity price prediction, and price difference prediction between day-ahead electricity price and real-time electricity price. It can not only locate the adjustment direction of the power curve more accurately, but also establish a more robust optimal declaration curve by comprehensively considering the accuracy of prediction models of different factors. By directly predicting the price difference instead of simple subtraction operation, the present invention successfully avoids the problem of introducing randomness, thereby improving the credibility and stability of the prediction. In addition, the model pays more attention to risk control, and through the careful adjustment of the adjustment ratio, it ensures that the uncertainty of different prediction models is fully considered in power trading decisions, making the overall model more adaptable and reliable. Therefore, the present invention overcomes the limitations of the prior art and provides a more advanced and reliable solution for power trading decisions. Summary of the invention
[0005] The present invention aims to solve the problems in the background technology. A power curve optimization method for the spot market based on artificial intelligence for new energy stations is adopted. By predicting the price difference between the day-ahead clearing electricity price and the real-time clearing electricity price, and combining multiple factors for comprehensive consideration, an optimal power curve declaration model is constructed. Compared with the prior art, the present invention not only emphasizes intelligent adjustment, but also pays attention to risk control, comprehensively considers the uncertainty of different prediction models, so as to optimize the electricity energy and electricity charges to the greatest extent. The method comprises the following steps:
[0006] S1, obtain data from multiple data sources;
[0007] S2, process the data and establish a data feature set;
[0008] S3, select important features;
[0009] S4, establish a day-ahead clearing electricity price prediction model based on machine learning to predict the day-ahead clearing electricity price;
[0010] S5, calculating the historical price difference, the method for calculating the historical price difference comprises the following steps:
[0011] Price difference = day-ahead clearing price - real-time clearing price
[0012] Adjustment direction: if the price difference is less than or equal to 0, the power curve is adjusted downward; if the price difference is greater than 0, the power curve is adjusted upward;
[0013] S6, establish multiple models to predict price differences;
[0014] S7, calculating the adjustment direction and adjustment probability;
[0015] S8, upward adjustment of the proportional grid and downward adjustment of the proportional grid calculation;
[0016] S9, calculating the regulation ratio at a certain moment according to the day-ahead predicted electricity price and the upward regulation ratio grid and the downward regulation ratio grid;
[0017] S10, dividing the price segments and updating the adjustment direction and adjustment probability;
[0018] S11, based on the adjustment ratio t and upward adjustment probability p 向上调节 , downward adjustment probability p 向下调节 , Q before adjustment 调整前 Power forecast and installed capacity calculation correction adjustment ratio t';
[0019] The specific steps are as follows:
[0020] Q 调整前 <20% installed capacity, t'=t*(1+|p 向上调节 -p 向下调节 );
[0021] 20% installed capacity <=<60% installed capacity, t=t;
[0022] Q 调整前 >=60% installed capacity, t'=t*(1-|p 向上调节 -p 向下调节 |);
[0023] S12, based on the above-mentioned correction adjustment ratio t', adjustment direction, and power prediction Q before adjustment 调整前 , calculate the adjusted Q 调整后 The power prediction is calculated as follows: Q 调整后 =t'*Q, if Q 调整后 <0, then; Q 调整后 =0, if Q 调整后 > installed capacity, then Q 调整后 =Installed capacity.
[0024] Furthermore, the data from multiple data sources obtained in step S1 include: one or more combinations of market supply and demand data, market settlement data, market contract data, market electricity price data, weather forecast data, coal price data, policy data, and new energy site power forecast data.
[0025] Furthermore, the market supply and demand data include the provincial load of the day-ahead market of the trading center, the predicted output of new energy, the output of non-marketized units, the total output of the province, the transmission interconnection line plan, the planned output of hydropower, the maintenance capacity, and the data of the units that must be started or shut down;
[0026] The market settlement data includes the day-ahead settlement data and real-time settlement data of the trading center;
[0027] The market contract data includes the medium- and long-term data of the trading center;
[0028] The market electricity price data includes the day-ahead clearing electricity price and the real-time clearing electricity price of the trading center;
[0029] The weather forecast data includes temperature, air pressure, wind speed, light, and humidity meteorological data;
[0030] The coal price data includes thermal coal data;
[0031] The policy data include policy data on clearing the upper and lower limits of electricity prices in the trading provinces;
[0032] The new energy station power prediction data includes photovoltaic or wind power power prediction curves.
[0033] Furthermore, the step S2 of establishing the feature set includes:
[0034] Statistics of each feature in the daily and hourly windows, including maximum, minimum, mean, median, upper quartile, lower quartile, and variance;
[0035] Statistical history lag characteristics:
[0036] The electricity price is also affected by the historical electricity price and the historical values of other input characteristics, so it is necessary to consider the historical lagged data. If D-day is the target day, the historical data of D-1, D-2, D-3, D-7, and D-14 should be considered;
[0037] Statistical time-related features:
[0038] Year, quarter, month, ten-day period, week, day of the month, day of the week, hour, and minute information;
[0039] Whether it is the beginning of the year, the end of the year, the end of the month, the beginning of the month, the weekend, a holiday, or the beginning of a quarter;
[0040] Information on the length of time since the beginning of the year, the length of time since the weekend, and the length of time since holidays.
[0041] Furthermore, the step S3 of selecting important features includes the following steps:
[0042] According to the correlation between features, redundant features with correlation less than 0.2 are removed;
[0043] Use the feature_importance method of the xgboost algorithm to select important features. XGBoost is an ensemble learning method that builds a powerful model by combining multiple weak learners (usually decision trees). It is trained in a gradient boosting manner, gradually improving the performance of the model through iteration, where the feature_importance method is used to measure the contribution of each feature in the model to the prediction target. This importance is relative and is calculated by analyzing the use of features in the model during the training process. Feature importance can be calculated in two main ways: one is the number of times the feature is selected during the splitting process of the tree, that is, the number of splits; the other is the average value of the information gain brought by the feature at each split. The specific process of selecting important features is as follows:
[0044] Model training: Use the XGBoost algorithm to train the data and build an integrated model.
[0045] Feature Importance Calculation: After training is completed, the importance score of each feature is calculated through the feature_importance method.
[0046] Importance sorting: The higher the value of feature importance, the greater the impact of the feature on the model's prediction. By sorting the features by importance, the most informative features can be identified.
[0047] Feature selection: Based on the ranking of feature importance, the most important features can be selected to be retained, thereby reducing the dimensionality and improving the interpretability and generalization ability of the model.
[0048] Furthermore, the establishment of the day-ahead clearing electricity price prediction model based on machine learning in step S4 includes the following steps:
[0049] Data preparation: Select the data from multiple data sources in step S1 as the data set of the model, select step S2 to create a feature set, divide the feature set into training set and test set data, and ensure that the data format is suitable for XGBoost, random forest and LSTM;
[0050] Divide the data set: Divide the training set into multiple subsets and use cross-validation to ensure that each subset has enough samples to represent the overall data distribution;
[0051] Feature engineering: The feature selection method of step S3 is used to screen features for each subset, and the top 50 features in terms of importance are selected as the input of the model;
[0052] Base model training: Use each subset to train the XGBoost model and adjust the hyperparameters to improve performance; use the same subset to train the random forest model and adjust the hyperparameters to improve performance; use the same subset to train the LSTM model and ensure that parameters such as data sequence length are compatible with the model;
[0053] Base model prediction: Use the test set data to make predictions through the trained XGBoost, random forest, and LSTM models;
[0054] Construct meta-features: The prediction results of each base model are used as new meta-features to form a new training set. Each sample in the new training set corresponds to the prediction results of the three models: XGBoost, Random Forest, and LSTM.
[0055] Metamodel training: Use the linear model as the metamodel and train it using the training set containing the prediction results of the base model and the corresponding target variables;
[0056] Stacking model prediction: Use the prediction results of the base model of the test set as meta-features and input them into the trained meta-model to obtain the final day-ahead clearing electricity price prediction model based on machine learning;
[0057] Model evaluation: Use evaluation indicators (such as root mean square error, mean absolute error, etc.) to evaluate the performance of the day-ahead clearing electricity price forecasting model based on machine learning to verify the forecasting ability of the model;
[0058] Based on the day-ahead clearing electricity price forecast results, support is provided for price difference forecast and adjustment ratio calculation.
[0059] Furthermore, the establishment of multiple models for price difference prediction in step S6 includes: establishing a clustering model, establishing a classification model, establishing a regression model and a combination model.
[0060] Furthermore, the establishment of the clustering model includes the following steps:
[0061] Data preparation: Select the data from multiple data sources in step S1 as the feature set of the model, standardize the feature set to standardize the features of different scales to the same scale, and divide the feature set into a training set and a test set;
[0062] Feature engineering: Use the feature selection method in step S3 and select the top 19 features in terms of feature importance as the input of the clustering model;
[0063] Model building: Establish a K-means clustering model. K-means is an unsupervised learning clustering algorithm. Its goal is to divide the data set into K clusters so that the data points within the cluster have high similarity, while the similarity between clusters is low. The algorithm is implemented by iteratively optimizing the center position of the cluster. The hyperparameter k value is calculated by the Gap statistic, and the K value with the largest Gap statistic is selected as the hyperparameter of the model.
[0064] Cluster adjustment direction and cluster adjustment probability: Each class contains data of adjustment directions 1 and 0. The proportion of 1 and 0 in each class is counted. If the proportion of 1 is greater than the proportion of 0, the adjustment direction of the class is 1, otherwise the adjustment direction of the class is 0. The proportion of upward adjustment probability is 1, and the proportion of downward adjustment probability is 0.
[0065] Model prediction: Use the test set to input into the trained K-means model to obtain the test results;
[0066] Model evaluation: Convert the clustering problem into a classification problem and use evaluation indicators to evaluate the performance of the K-means model to verify the predictive ability of the model;
[0067] Furthermore, establishing a classification model includes the following steps:
[0068] Data preparation: Select the data from multiple data sources in step S1 as the data set of the model, and add the day-ahead clearing electricity price predicted in step S4 to the data set. Select step S2 to establish a feature set, divide the feature set into training set and test set data, and ensure that the data format is suitable for LightGBM;
[0069] Feature engineering: The feature selection method in step S3 is used to screen the features of the training set and the test set, and the top 50 features in terms of importance are selected as the input of the model;
[0070] Model building: Building the LightGBM classification model, which uses a gradient boosting algorithm based on decision tree integration. Each round of iteration generates a new tree, and the final model is obtained through weighted summation.
[0071] The LightGBM classification model is constructed by setting the hyperparameter learning rate, number of leaf nodes, maximum depth of the tree, minimum number of leaf node samples, and feature sampling ratio. The Bayesian optimization algorithm is used to optimize the hyperparameters of the LightGBM model. The optimal parameter combination is found through continuous iteration, and the optimized optimal parameters are input into the LightGBM model to construct an adjustment direction classification prediction model.
[0072] Model prediction: Use the test set to input into the trained LightGBM model to get the test results;
[0073] Classification adjustment direction and classification adjustment probability: When dealing with binary classification problems, LightGBM usually outputs a probability value, which indicates the probability that a sample belongs to the positive class, that is, the probability that the adjustment direction is 1, and the probability that the adjustment direction is 0 is 1 minus the probability that the adjustment direction is 1. The prediction result of LightGBM is the classification adjustment direction;
[0074] Model evaluation: Use evaluation metrics to evaluate the performance of the LightGBM model to verify the predictive ability of the model;
[0075] Furthermore, building a regression model includes the following steps:
[0076] Data preparation: Select the data from multiple data sources in step S1 as the data set of the model, and add the day-ahead clearing electricity price predicted in step S4 to the data set. Select step S2 to create a feature set, divide the feature set into training set and test set data, and ensure that the data format is suitable for XGBoost.
[0077] Feature engineering: The feature selection method of step S3 is used to screen the features of the training set and the test set, and the top 50 features in terms of importance are selected as the input of the model;
[0078] Model building: Establish an XGBoost regression model by setting the hyperparameter learning rate, the maximum depth of the tree, the minimum sample weight of the leaf node, and the feature sampling ratio. Apply the Bayesian optimization algorithm to optimize the hyperparameters of the XGBoost model, find the optimal parameter combination through continuous iteration, input the optimized optimal parameters into the XGBoost model, and build a price difference regression prediction model.
[0079] Model prediction: Use the test set to input into the trained XGBoost model to get the test results;
[0080] Regression adjustment direction and regression adjustment probability: The regression adjustment direction is determined by judging whether the predicted price difference is positive or negative. If the price difference is less than or equal to 0, the regression adjustment direction is 0. If the price difference is greater than , the regression adjustment direction is 1. The calculation of the adjustment probability is calculated by introducing the sigmoid function. The expression of the sigmoid function is:
[0081] Where x is the predicted price difference, the price difference is used as the input of the sigmoid function, and the price difference is mapped to the range of (0,1), which represents the regression adjustment probability;
[0082] Model evaluation: Convert the regression problem into a classification problem and use evaluation indicators (accuracy, precision, recall, ROC curve, AUC) to evaluate the performance of the XGBoost model to verify the predictive ability of the model.
[0083] Furthermore, the adjustment probability calculation method in step S7 is:
[0084]
[0085]
[0086] The adjustment direction and adjustment probability of the three models are as follows:
[0087]
[0088] Of which: Acc 1 、Acc 2 、Acc 3 are the accuracy of clustering model, classification model and regression model respectively; w 1 is the weight in the clustering model, w 2 is the weight in the classification model, w 3 is the weight in the regression model, p 1 , p 2 , p 3 is the upward adjustment probability of the clustering model, classification model, and regression model, q 1 ,q 2 ,q 3 Downward adjusted probabilities for clustering, classification, and regression models;
[0089] Furthermore, the steps of calculating the upward adjustment ratio grid and the downward adjustment ratio grid in step S8 are as follows:
[0090] Get the day-ahead electricity price and real-time electricity price data 120 days before the forecast date and the forecast power before adjustment
[0091] Q 调整前 , the data are one point every 15 minutes, 96 moments in a day;
[0092] Extract all data with adjustment direction of 1 within 120 days as data1, and extract all data with adjustment direction of 0 within 120 days as data2;
[0093] The day-ahead clearing electricity prices in data1 and data2 are divided into buckets, and the bucket width is set to 100 yuan / MWh;
[0094] Data1 and data2 are grouped and counted according to 96 moments and buckets, forming a grid with 96 moments as buckets. Each grid contains 0 or more day-ahead clearing prices, real-time clearing prices, and the predicted power Q before adjustment. 调整前 data;
[0095] The day-ahead clearing price, real-time clearing price and Q 调整前 Normalization is performed to form the normalized day-ahead clearing electricity price P′ 日前 , normalized real-time clearing electricity price P′ 实时 And the normalized power prediction before regulation Q′ 调整前 .
[0096] Based on the trading rules, determine the overall profit of the power forecast before normalization at the current moment
[0097] R′ 调整前 and the normalized adjusted power predicted overall gain R 调整后 , the specific formula is as follows:
[0098] R′ 调整前 =Q′ 中长期 P′ 中长期 +(Q′ 调整前 -Q′ 中长期 )P′ 日前 +(Q′ 实时 -Q′ 调整前 )P′ 实时
[0099] R′ 调整后 =Q′ 中长期 P′ 中长期 +(Q′ 调整后 -Q′ 中长期 )P′ 日前 +(Q′ 实时 -Q′ 调整后 )P′ 实时
[0100] where Q′ 中长期 is the normalized medium and long term electricity; P′ 中长期 is the normalized medium- and long-term electricity price; Q′ 实时is the normalized actual online power consumption;
[0101] Establish the objective function f = R' 调整后 -R′ 调整前 The target result of the objective function f is used to determine the optimal adjustment ratio of the declared adjustment at the current moment. According to the benefit function, it can be calculated
[0102] f=(Q′ 调整后 -Q′ 调整前 )*(P′ 日前 -P′ 实时 ), where Q′ 调整后 =t*Q′ 调整前 , t is the adjustment ratio, and further we can get f=(t-1)*Q′ 调整前 *(P′ 日前 -P′ 实时 ), it can be seen that the benefit is a function of the adjustment ratio;
[0103] In the data set data1, the range of the adjustment ratio t is set to [1, 2.5], and the interval of the adjustment ratio is 0.1. Specifically, for each price difference value, 16 revenue values under different adjustment ratios can be calculated. Finally, the median of all revenues corresponding to all price differences is obtained, and the adjustment ratio corresponding to the median is used as the upward adjustment ratio under the bucket at that moment. If there is no data in the bucket at that moment, the adjustment ratio is set to 1;
[0104] In the data set data2, the value range of the adjustment ratio t is set to [0,1), and the interval of the adjustment ratio is 0.1. Specifically, for each price difference value, 10 revenue values under different adjustment ratios can be calculated. Finally, the median of all revenues corresponding to all price differences is obtained, and the adjustment ratio corresponding to the median is used as the downward adjustment ratio under the bucket at that moment. If there is no data in the bucket at that moment, the adjustment ratio is set to 1. ;
[0105] Thus, an upward adjustment ratio grid and a downward adjustment ratio grid are obtained;
[0106] Furthermore, the method of dividing the price segments and updating the adjustment direction and adjustment probability in step S10 includes:
[0107] The day-ahead forecast electricity price is divided into three categories: the lower limit of the provincial clearing price, the upper limit of the provincial clearing price, and other electricity prices that do not belong to the two. If the day-ahead forecast electricity price is the lower limit of the provincial clearing price, the adjustment direction is set to downward adjustment, and the adjustment probability is set to 1. If the day-ahead forecast electricity price is the upper limit of the provincial clearing price, the adjustment direction is set to upward adjustment, and the adjustment probability is set to 1. If the day-ahead clearing electricity price is other electricity prices, no correction is made to the adjustment direction and adjustment probability.
[0108] The document content enhanced retrieval system and method for multimodal information fusion provided by the present invention solves the problem of predicting the size and direction of the price difference between the day-ahead price and the real-time price, thereby determining the adjustment direction of the power prediction curve of the new energy station in the day-ahead market.
[0109] 1. Based on the above, after determining the adjustment direction of the power curve, the adjustment risk caused by the prediction accuracy and uncertainty is taken into account, and the problem of adjustment ratio is solved.
[0110] 2. The electricity trading market is a market that changes at any time according to trading rules and weather conditions. Therefore, the power curve adjustment ratio is not fixed, which solves the problem of adaptive adjustment ratio.
[0111] 3. At the same time, intelligent methods are used to reduce the risks and labor costs of spot transactions. BRIEF DESCRIPTION OF THE DRAWINGS
[0112] Figure 1 This is a flow chart of the power curve optimization method for the spot market of a new energy station based on AI proposed by the present invention;
[0113] Figure 2 This is a flow chart of the day-ahead clearing electricity price prediction of the power curve optimization method for the spot market based on AI for new energy stations proposed by the present invention;
[0114] Figure 3 A flow chart of the clustering model, classification model, and regression model for predicting price differences of the power curve optimization method for the spot market based on AI for new energy stations proposed by the present invention;
[0115] Figure 4 This is a grid flow chart of the upward and downward adjustment ratios of the power curve optimization method for the spot market based on AI for new energy stations proposed in the present invention. DETAILED DESCRIPTION
[0116] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.
[0117] The main focus of the present invention is the price difference between the day-ahead electricity price and the real-time electricity price and the optimal adjustment ratio. The direction of the price difference determines whether the adjustment is correct, and the adjustment ratio prevents economic losses caused by forecast deviations.
[0118] like Figure 1 The specific steps shown are:
[0119] 1. Basic data
[0120] Supply and demand data: the day-ahead market's provincial load, new energy output, non-marketized unit output, provincial total output, external transmission lines, hydropower output, maintenance capacity, data on must-start and must-stop units, day-ahead clearing electricity prices, and real-time clearing electricity prices;
[0121] Meteorological data: obtain temperature, air pressure, wind speed, light, and humidity meteorological data;
[0122] Coal price data: thermal coal data
[0123] Policy data: such as upper and lower limits of electricity prices, etc.
[0124] 2. Process the above data:
[0125] 2.1 Feature Statistics
[0126] Count the statistical values of each digital feature under different windows, such as maximum value, minimum value, variance, mean, median, etc. Currently, the windows are selected as day and hour, that is, the window lengths are 96 and 4 respectively.
[0127] 2.2 Historical hysteresis characteristics
[0128] The electricity price is also affected by the historical electricity prices and the historical values of other input characteristics, so it is necessary to consider the historical lagged data and the historical data of D-1, D-2, D-3, D-7 and D-14.
[0129] 2.3 Time-related features
[0130] Considering that electricity prices change according to time and have time series characteristics, it is necessary to consider time series related characteristics. Time series related characteristics include time-based characteristics, Boolean characteristics, and time difference characteristics.
[0131] Time-based features: year, quarter, month, ten-day period, week, day of the month, day of the week, hour, minute, etc.
[0132] Boolean features: whether it is the beginning of the year, the end of the year, the end of the month, the beginning of the month, the weekend, a holiday, the beginning of a quarter, etc.
[0133] Time difference characteristics: information such as the length of time since the beginning of the year, the length of time since the weekend, and the length of time since holidays.
[0134] 3. Feature Selection
[0135] 1) According to the correlation between features, remove redundant features with correlation less than 0.2
[0136] 2) Based on 1), the feature_importance method of the xgboost algorithm is used to select important features;
[0137] XGBoost is an ensemble learning method that builds a powerful model by combining multiple weak learners (usually decision trees). It is trained in a gradient boosting manner, gradually improving the performance of the model through iteration, where the feature_importance method is used to measure the contribution of each feature in the model to the prediction target. This importance is relative and is calculated by analyzing the use of features in the model during the training process. Feature importance can be calculated in two main ways: one is the number of times the feature is selected during the tree splitting process, that is, the number of splits; the other is the average value of the information gain brought by the feature at each split. The specific process of selecting important features is as follows:
[0138] Model training: Use the XGBoost algorithm to train the data and build an integrated model.
[0139] Feature Importance Calculation: After training is completed, the importance score of each feature is calculated through the feature_importance method.
[0140] Importance sorting: The higher the value of feature importance, the greater the impact of the feature on the model's prediction. By sorting the features by importance, the most informative features can be identified.
[0141] Feature selection: Based on the ranking of feature importance, the most important features can be selected to be retained, thereby reducing the dimensionality and improving the interpretability and generalization ability of the model.
[0142] 4. Establish electricity price prediction model:
[0143] A stacking fusion model based on xgboost, lightgbm and catboost was established to predict the day-ahead clearing electricity price.
[0144] like Figure 2 The day-ahead clearing electricity price forecasting model shown includes the following steps:
[0145] Data preparation: Select the data from multiple data sources in step S1 as the data set of the model, select step S2 to create a feature set, divide the feature set into training set and test set data, and ensure that the data format is suitable for XGBoost, random forest and LSTM;
[0146] Divide the data set: Divide the training set into multiple subsets and use cross-validation to ensure that each subset has enough samples to represent the overall data distribution;
[0147] Feature engineering: that is, selecting important features. Using the method in step S3, perform feature screening on each subset and select the top 50 features as the input of the model.
[0148] Base model training: Use each subset to train the XGBoost model and adjust the hyperparameters to improve performance; use the same subset to train the random forest model and adjust the hyperparameters to improve performance; use the same subset to train the LSTM model and ensure that parameters such as the data sequence length are adapted to the model.
[0149] Base model prediction: Use the test set data to make predictions through the trained XGBoost, random forest, and LSTM models;
[0150] Construct meta-features: Take the prediction results of each base model as new features (meta-features) to form a new training set. Each sample in the new training set corresponds to the prediction results of the three models: XGBoost, Random Forest, and LSTM.
[0151] Metamodel training: Use the linear model as the metamodel, and use the new training set (including the prediction results of the base model) and the corresponding target variables for training;
[0152] Stacking model prediction: Use the base model prediction results of the test set as meta-features and input them into the trained meta-model to obtain the final Stacking model prediction results;
[0153] Model evaluation: Use evaluation indicators (such as root mean square error, mean absolute error, etc.) to evaluate the performance of the stacking model to verify the predictive ability of the model;
[0154] Based on the day-ahead clearing electricity price forecast results, support is provided for price difference forecast and adjustment ratio calculation.
[0155] 5. Calculate historical price difference
[0156] Price difference = day-ahead clearing price - real-time clearing price
[0157] Spread direction: when the spread is less than or equal to 0, adjust downward; when the spread is greater than 0, adjust upward. The reason why the spread is adjusted downward when the spread is equal to 0 is to convert the three-classification problem into a two-classification problem, which will greatly improve the prediction accuracy. The reason for this operation is that when the spread is 0, the adjustment direction does not affect the profit.
[0158] 6. If Figure 3 Spread forecast shown
[0159] 6.1 Clustering Model
[0160] By building a price difference direction clustering model, we can derive the probability of upward and downward adjustments. The specific steps are as follows:
[0161] 1) The data set uses the basic data in step 1 as the feature set of the clustering model.
[0162] 2) Data cleaning: check and process missing values, outliers, etc.
[0163] 3) Not all of the above basic data are required to participate in the establishment of the clustering model. Therefore, the xgboost feature importance method is used to select the top 19 features as the input of the model.
[0164] 4) Feature standardization: Standardize features of different scales to the same scale
[0165] 5) Establish k-means clustering model
[0166] 6) The clustering price difference direction and clustering adjustment probability are obtained through the clustering model (referring to the price difference direction and adjustment probability obtained through the clustering model)
[0167] 6.2 Classification Model
[0168] By building a price difference direction classification model, we can also derive the probability of upward and downward adjustments. The specific steps are as follows:
[0169] 1) The data set uses the data processed in step 2 above as the feature set
[0170] 2) Use the feature_importance method of the xgboost algorithm to select important features, and select the top 50 features of importance as the input of the model
[0171] 3) Establish MLP (Multi-layer Perceptron) classification model
[0172] 4) The classification price difference direction and classification adjustment probability are obtained through the classification model (referring to the price difference direction and adjustment probability obtained through the classification model)
[0173] 6.3 Regression Model
[0174] A regression model of the price difference is established. After the price difference is obtained, the regression output price difference is converted into a probability value through the sigmoid function. The probability of upward and downward adjustment can also be obtained. The specific steps are as follows:
[0175] 1) The data set uses the data processed in step 2 above as the feature set
[0176] 2) Use the feature_importance method of the xgboost algorithm to select important features, and select the top 44 features of importance as the input of the model
[0177] 3) Establish an Xgboost regression model based on Bayesian optimization
[0178] 4) The price difference is obtained through the regression model, and the direction of the regression price difference is further calculated (the price difference is adjusted upward if it is positive, and downward if it is negative)
[0179] 5) Use the price difference as the input of the sigmoid function to calculate the regression adjustment probability.
[0180] Specifically:
[0181] Establishing multiple models for price difference prediction includes: establishing clustering model, classification model, regression model and combination model;
[0182] Described establishment cluster model comprises the following steps:
[0183] Data preparation: Select the data from multiple data sources in step S1 as the feature set of the model, standardize the feature set to standardize the features of different scales to the same scale, and divide the feature set into a training set and a test set;
[0184] Feature engineering: Use the feature selection method in step S3 and select the top 19 features in terms of feature importance as the input of the clustering model;
[0185] Model building: Establish a K-means clustering model. K-means is an unsupervised learning clustering algorithm. Its goal is to divide the data set into K clusters so that the data points within the cluster are highly similar, while the similarity between clusters is low. The algorithm is implemented by iteratively optimizing the center position of the cluster. The hyperparameter k value is calculated by the Gap statistic, and the K value with the largest Gap statistic is selected as the hyperparameter of the model;
[0186] Cluster adjustment direction and cluster adjustment probability: Each class contains data of adjustment directions 1 and 0. The proportion of 1 and 0 in each class is counted. If the proportion of 1 is greater than the proportion of 0, the adjustment direction of the class is 1, otherwise the adjustment direction of the class is 0. The proportion of upward adjustment probability is 1, and the proportion of downward adjustment probability is 0.
[0187] Model prediction: Use the test set to input into the trained K-means model to obtain the test results;
[0188] Model evaluation: Convert the clustering problem into a classification problem, and use evaluation indicators (accuracy, precision, recall, ROC curve, AUC) to evaluate the performance of the K-means model to verify the predictive ability of the model;
[0189] Building a classification model involves the following steps:
[0190] Data preparation: Select the data from multiple data sources in step S1 as the data set of the model, and add the day-ahead clearing electricity price predicted in step S4 to the data set. Select step S2 to establish a feature set, divide the feature set into training set and test set data, and ensure that the data format is suitable for LightGBM;
[0191] Feature engineering: The feature selection method of step S3 is used to screen the features of the training set and the test set, and the top 50 features in terms of importance are selected as the input of the model;
[0192] Model building: Building the LightGBM classification model, which uses a gradient boosting algorithm based on decision tree integration. Each round of iteration generates a new tree, and the final model is obtained through weighted summation.
[0193] The LightGBM classification model is constructed by setting the hyperparameter learning rate, number of leaf nodes, maximum depth of the tree, minimum number of leaf node samples, and feature sampling ratio. The Bayesian optimization algorithm is used to optimize the hyperparameters of the LightGBM model. The optimal parameter combination is found through continuous iteration, and the optimized optimal parameters are input into the LightGBM model to construct an adjustment direction classification prediction model.
[0194] Model prediction: Use the test set to input into the trained LightGBM model to get the test results;
[0195] Classification adjustment direction and classification adjustment probability: When dealing with binary classification problems, LightGBM usually outputs a probability value, which indicates the probability that a sample belongs to the positive class, that is, the probability that the adjustment direction is 1, and the probability that the adjustment direction is 0 is 1 minus the probability that the adjustment direction is 1. The prediction result of LightGBM is the classification adjustment direction;
[0196] Model evaluation: Use evaluation indicators (accuracy, precision, recall, ROC curve, AUC) to evaluate the performance of the LightGBM model to verify the predictive ability of the model;
[0197] Building a regression model involves the following steps:
[0198] Data preparation: Select the data from multiple data sources in step S1 as the data set of the model, and add the day-ahead clearing electricity price predicted in step S4 to the data set. Select step S2 to create a feature set, divide the feature set into training set and test set data, and ensure that the data format is suitable for XGBoost.
[0199] Feature engineering: The feature selection method of step S3 is used to screen the features of the training set and the test set, and the top 50 features in terms of importance are selected as the input of the model;
[0200] Model building: Establish an XGBoost regression model by setting the hyperparameter learning rate, the maximum depth of the tree, the minimum sample weight of the leaf node, and the feature sampling ratio. Apply the Bayesian optimization algorithm to optimize the hyperparameters of the XGBoost model, find the optimal parameter combination through continuous iteration, input the optimized optimal parameters into the XGBoost model, and build a price difference regression prediction model.
[0201] Model prediction: Use the test set to input into the trained XGBoost model to get the test results;
[0202] Regression adjustment direction and regression adjustment probability: The regression adjustment direction is determined by judging whether the predicted price difference is positive or negative. If the price difference is less than or equal to 0, the regression adjustment direction is 0. If the price difference is greater than , the regression adjustment direction is 1. The calculation of the adjustment probability is calculated by introducing the sigmoid function. The expression of the sigmoid function is:
[0203] Where x is the predicted price difference, the price difference is used as the input of the sigmoid function, and the price difference is mapped to the range of (0,1), which represents the regression adjustment probability;
[0204] Model evaluation: Convert the regression problem into a classification problem and use evaluation indicators (accuracy, precision, recall, ROC curve, AUC) to evaluate the performance of the XGBoost model to verify the predictive ability of the model.
[0205] 7. If Figure 4 The calculations shown here scale the mesh up and scale the mesh down:
[0206] Get the day-ahead electricity price and real-time electricity price data 120 days before the forecast date and the forecast power before adjustment
[0207] Q 调整前 The data is one point every 15 minutes, 96 times a day.
[0208] Extract all data with adjustment direction of 1 within 120 days as data1, and extract all data with adjustment direction of 0 within 120 days as data2;
[0209] The day-ahead clearing electricity prices in data1 and data2 are divided into buckets, and the bucket width is set to 100 yuan / MWh;
[0210] Data1 and data2 are grouped and counted according to 96 moments and buckets, forming a grid with 96 moments as buckets. Each grid contains 0 or more day-ahead clearing prices, real-time clearing prices, and the predicted power Q before adjustment. 调整前 data;
[0211] The day-ahead clearing price, real-time clearing price and Q 调整前 Normalization is performed to form the normalized day-ahead clearing electricity price P′ 日前 , normalized real-time clearing electricity price P′ 实时 And the normalized power prediction before regulation Q′ 调整前 ;
[0212] Based on the trading rules, determine the overall profit of the power forecast before normalization at the current moment
[0213] R′ 调整前 and the normalized adjusted power prediction overall benefit R′ 调整后 , the specific formula is as follows:
[0214] R′ 调整前 =Q′ 中长期 P′ 中长期 +(Q′ 调整前 -Q′ 中长期 )P′ 日前 +(Q′ 实时 -Q′ 调整前 )P′ 实时
[0215] R′ 调整后 =Q′ 中长期 P′ 中长期 +(Q′ 调整后 -Q′ 中长期 )P′ 日前 +(Q′ 实时 -Q′ 调整后 )P′ 实时
[0216] where Q′ 中长期 is the normalized medium and long term electricity; P′ 中长期 is the normalized medium- and long-term electricity price; Q′ 实时 is the normalized actual online power consumption.
[0217] Establish the objective function f = R' 调整后 -R′ 调整前 The target result of the objective function f is used to determine the optimal adjustment ratio of the declared adjustment at the current moment. According to the benefit function, it can be calculated
[0218] f=(Q′ 调整后 -Q′ 调整前 )*(P′ 日前 -P′ 实时 ) where Q′ 调整后 =t*Q′ 调整前 , t is the adjustment ratio, and further we can get f=(t-1)*Q′ 调整前*(P′ 日前 -P′ 实时 ) It can be seen that the return is a function of the adjustment ratio;
[0219] In the data set data1, the range of the adjustment ratio t is set to [1, 2.5], and the interval of the adjustment ratio is 0.1. Specifically, for each price difference value, 16 revenue values under different adjustment ratios can be calculated. Finally, the median of all revenues corresponding to all price differences is obtained, and the adjustment ratio corresponding to the median is used as the upward adjustment ratio under the bucket at that moment. If there is no data in the bucket at that moment, the adjustment ratio is set to 1;
[0220] In the data set data2, the range of the adjustment ratio t is set to [0, 1), and the interval of the adjustment ratio is 0.1. Specifically, for each price difference value, 10 revenue values under different adjustment ratios can be calculated. Finally, the median of all revenues corresponding to all price differences is obtained, and the adjustment ratio corresponding to the median is used as the downward adjustment ratio under the bucket at that moment. If there is no data in the bucket at that moment, the adjustment ratio is set to 1;
[0221] Thus, an upward adjustment ratio grid and a downward adjustment ratio grid are obtained;
[0222] 8. Calculation method of adjustment probability:
[0223] The prediction accuracy of clustering, classification, and regression models is calculated through historical data, which are
[0224] Acc 1 、Acc 2 、Acc 3 , calculate the weight of each model in the comprehensive price difference direction determination, respectively:
[0225]
[0226] The price difference direction and adjustment probability of the three models are as follows:
[0227]
[0228] Where: p 1 , p 2 , p 3 is the upward adjusted probability for clustering, classification, and regression,
[0229] q 1 ,q 2 ,q 3The probability of downward adjustment of clustering, classification, and regression. The above is the adjustment probability obtained through clustering, classification, and regression, and a new adjustment probability is calculated by combining them; 9. Adjustment ratio correction
[0230] 1) Calculate the adjustment ratio t through the forecast value of the day-ahead clearing electricity price and the upward adjustment ratio grid and the downward adjustment ratio grid;
[0231] 2 ), based on the adjustment ratio t, upward adjustment probability p 向上调节 , downward adjustment probability p 向下调节 , Q before adjustment 调整前 Power forecast and installed capacity calculation correction adjustment ratio t';
[0232] The specific steps are as follows:
[0233] Q 调整前 <20% installed capacity, t'=t*(1+|p 向上调节 -p 向下调节 |);
[0234] 20% installed capacity <=<60% installed capacity, t'=t;
[0235] Q 调整前 >=60% installed capacity, t'=t*(1-|p 向上调节 -p 向下调节 |);
[0236] 10. Based on the above correction adjustment ratio t', adjustment direction, and power prediction Q before adjustment 调整前 , calculate the adjusted Q 调整后 The power forecast is reported to the dispatcher. The calculation method is: Q 调整后 =t'*Q, if Q 调整后 <0, then; Q 调整后 =0, if Q 调整后 > installed capacity, then Q 调整后 =Installed capacity.
[0237] It should be understood that the embodiments herein are only used to illustrate the present invention and are not used to limit the scope of the present invention. Various changes made by ordinary technicians in this field based on the above concepts without creative work are all within the protection scope of the present invention.
Claims
1. The power curve optimization method for new energy stations based on AI for the spot market is characterized by: The method comprises the following steps: S1, obtain data from multiple data sources; S2, process the data and establish a data feature set; S3, select important features; S4, establish a day-ahead clearing electricity price prediction model based on machine learning to predict the day-ahead clearing electricity price; S5, calculating the historical price difference, the method for calculating the historical price difference comprises the following steps: Price difference = day-ahead clearing price - real-time clearing price Adjustment direction: if the price difference is less than or equal to 0, the power curve is adjusted downward; if the price difference is greater than 0, the power curve is adjusted upward; S6, establish multiple models to predict price differences; S7, calculating the adjustment direction and adjustment probability; S8, upward adjustment of the proportional grid and downward adjustment of the proportional grid calculation; S9, calculating the regulation ratio at a certain moment according to the day-ahead predicted electricity price and the upward regulation ratio grid and the downward regulation ratio grid; S10, dividing the price segments and updating the adjustment direction and adjustment probability; S11, based on the adjustment ratio t and upward adjustment probability p 向上调节 , downward adjustment probability p 向下调节 , Q before adjustment 调整前 Power forecast and installed capacity calculation correction adjustment ratio t'; The specific steps are as follows: Q 调整前 <20% installed capacity, t'=t*(1+|p 向上调节 -p 向下调节 |); 20% installed capacity <= <60% installed capacity, t'=t; Q 调整前 >=60% installed capacity, t'=t*(1-|p 向上调节 -p 向下调节 |); S12, based on the above-mentioned correction adjustment ratio t', adjustment direction, and power prediction Q before adjustment 调整前 , calculate the adjusted Q 调整后 The power prediction is calculated as follows: Q 调整后 =t'*Q, if Q 调整后 <0, then; Q 调整后 =0, if Q 调整后 >=0 installed capacity, then Q 调整后 =Installed capacity.
2. The power curve optimization method for the spot market of a new energy station based on AI according to claim 1 is characterized in that: The data from multiple data sources obtained in step S1 include: one or more combinations of market supply and demand data, market settlement data, market contract data, market electricity price data, weather forecast data, coal price data, policy data, and new energy station power forecast data; The market supply and demand data include the provincial load of the day-ahead market of the trading center, the predicted output of new energy, the output of non-marketized units, the total output of the province, the transmission interconnection line plan, the planned output of hydropower, the maintenance capacity, and the data of the units that must be started or shut down; The market settlement data includes the day-ahead settlement data and real-time settlement data of the trading center; The market contract data includes the medium- and long-term data of the trading center; The market electricity price data includes the day-ahead clearing electricity price and the real-time clearing electricity price of the trading center; The weather forecast data includes temperature, air pressure, wind speed, light, and humidity meteorological data; The coal price data includes thermal coal data; The policy data include policy data on clearing the upper and lower limits of electricity prices in the trading provinces; The new energy station power prediction data includes photovoltaic or wind power power prediction curves.
3. The power curve optimization method for the spot market of a new energy station based on AI according to claim 1 is characterized in that: The step S2 of establishing a data feature set for data processing includes: Characteristic statistics: Count the statistical values of each digital feature in the day and hour window, and the digital features include maximum value, minimum value, mean value, median value, upper quartile, lower quartile, and variance; Historical hysteresis characteristics: The historical data of D-1, D-2, D-3, D-7 and D-14 are collected, with D-day being the target day; Timing-related features: time-based features, Boolean features, and time difference features.
4. The power curve optimization method for the spot market of a new energy station based on AI according to claim 1 is characterized in that: The step S3 of selecting important features comprises the following steps: According to the correlation between features, redundant features with correlation less than 0.2 are removed; Use the feature_importance method of the xgboost algorithm to select important features.
5. The power curve optimization method for the spot market of a new energy station based on AI according to claim 1 is characterized in that: The step S4 of the day-ahead clearing electricity price prediction model comprises the following steps: Data preparation: Select the data from multiple data sources in step S1 as the data set of the model, select step S2 to create a feature set, divide the feature set into training set and test set data, and ensure that the data format is suitable for XGBoost, random forest and LSTM; Divide the data set: Divide the training set into multiple subsets and use cross-validation to ensure that each subset has enough samples to represent the overall data distribution; Feature engineering: that is, selecting important features. Using the method in step S3, perform feature screening on each subset and select the top 50 features as the input of the model. Base model training: Use each subset to train the XGBoost model and adjust the hyperparameters to improve performance; use the same subset to train the random forest model and adjust the hyperparameters to improve performance; use the same subset to train the LSTM model and ensure that parameters such as data sequence length are compatible with the model; Base model prediction: Use the test set data to make predictions through the trained XGBoost, random forest, and LSTM models; Construct meta-features: Take the prediction results of each base model as new features (meta-features) to form a new training set. Each sample in the new training set corresponds to the prediction results of the three models: XGBoost, Random Forest, and LSTM. Metamodel training: Use the linear model as the metamodel, and use the new training set (including the prediction results of the base model) and the corresponding target variables for training; Stacking model prediction: Use the base model prediction results of the test set as meta-features and input them into the trained meta-model to obtain the final Stacking model prediction results; Model evaluation: Use evaluation indicators (such as root mean square error, mean absolute error, etc.) to evaluate the performance of the stacking model to verify the predictive ability of the model.
6. The power curve optimization method for the spot market of a new energy station based on AI according to claim 1 is characterized in that: The establishment of multiple models for price difference prediction in step S6 includes: establishing a clustering model, establishing a classification model, establishing a regression model and a combination model; Described establishment cluster model comprises the following steps: Data preparation: Select the data from multiple data sources in step S1 as the feature set of the model, standardize the feature set to standardize the features of different scales to the same scale, and divide the feature set into a training set and a test set; Feature engineering: Use the feature selection method in step S3 and select the top 19 features in terms of feature importance as the input of the clustering model; Model establishment: Establish a K-means clustering model, in which the hyperparameter k value is calculated by the Gap statistic, and the K value with the largest Gap statistic is selected as the hyperparameter of the model; Cluster adjustment direction and cluster adjustment probability: Count the proportion of 1 and 0 in each class. If the proportion of 1 is greater than the proportion of 0, the adjustment direction of the class is 1. Otherwise, the adjustment direction of the class is 0. The proportion of upward adjustment probability is 1, and the proportion of downward adjustment probability is 0. Model prediction: Use the test set to input into the trained K-means model to obtain the test results; Model evaluation: Convert the clustering problem into a classification problem and use evaluation indicators to evaluate the performance of the K-means model to verify the predictive ability of the model. The evaluation indicators include accuracy, precision, recall, ROC curve, and AUC. Building a classification model involves the following steps: Data preparation: Select the data from multiple data sources in step S1 as the data set of the model, and add the day-ahead clearing electricity price predicted in step S4 to the data set. Select step S2 to establish a feature set, divide the feature set into training set and test set data, and ensure that the data format is suitable for LightGBM; Feature engineering: The feature selection method of step S3 is used to screen the features of the training set and the test set, and the top 50 features in terms of importance are selected as the input of the model; Model establishment: Establish a LightGBM classification model by setting the hyperparameter learning rate, number of leaf nodes, maximum tree depth, minimum number of leaf node samples, and feature sampling ratio. Apply the Bayesian optimization algorithm to optimize the hyperparameters of the LightGBM model, input the optimized optimal parameters into the LightGBM model, and build an adjustment direction classification prediction model. Model prediction: Use the test set to input into the trained LightGBM model to get the test results; Classification adjustment direction and classification adjustment probability: The prediction result of LightGBM is the classification adjustment direction. The probability value output by LightGBM when dealing with binary classification problems is the classification adjustment probability. Model evaluation: Use evaluation indicators to evaluate the performance of the LightGBM model to verify the predictive ability of the model, where the evaluation indicators include accuracy, precision, recall, ROC curve, and AUC; Building a regression model involves the following steps: Data preparation: Select the data from multiple data sources in step S1 as the data set of the model, and add the day-ahead clearing electricity price predicted in step S4 to the data set. Select step S2 to create a feature set, divide the feature set into training set and test set data, and ensure that the data format is suitable for XGBoost. Feature engineering: The feature selection method of step S3 is used to screen the features of the training set and the test set, and the top 50 features in terms of importance are selected as the input of the model; Model building: The XGBoost regression model is constructed by setting the hyperparameter learning rate, the maximum depth of the tree, the minimum sample weight of the leaf node, and the feature sampling ratio. The Bayesian optimization algorithm is used to optimize the hyperparameters of the XGBoost model. The optimized optimal parameters are input into the XGBoost model to construct a price difference regression prediction model. Model prediction: Use the test set to input into the trained XGBoost model to get the test results; Regression adjustment direction and regression adjustment probability: The regression adjustment direction is determined by judging whether the predicted price difference is positive or negative. If the price difference is less than or equal to 0, the regression adjustment direction is 0. If the price difference is greater than , the regression adjustment direction is 1. The regression adjustment probability is calculated by introducing the sigmoid function. The expression of the sigmoid function is: Where x is the predicted price difference, the price difference is used as the input of the sigmoid function, and the price difference is mapped to the range of (0,1), which represents the regression adjustment probability; Model evaluation: Convert the regression problem into a classification problem and use evaluation indicators to evaluate the performance of the XGBoost model to verify the predictive ability of the model. The evaluation indicators include accuracy, precision, recall, ROC curve, and AUC.
7. The power curve optimization method for the spot market of a new energy station based on AI according to claim 1 is characterized in that: The adjustment probability calculation method in step S7 is: The adjustment direction and adjustment probability of the three models are as follows: Among them: Acc1, Acc2, Acc3 are the accuracy rates of the clustering model, classification model, and regression model respectively; w1 is the weight in the clustering model, w2 is the weight in the classification model, w3 is the weight in the regression model, p1, p2, p3 are the probabilities of upward adjustment of the clustering model, classification model, and regression model, q1, q2, q3 are the probabilities of downward adjustment of the clustering model, classification model, and regression model.
8. The power curve optimization method for the spot market of a new energy station based on AI according to claim 1 is characterized in that: The steps of calculating the upward adjustment ratio grid and the downward adjustment ratio grid in step S8 are as follows: Take the day-ahead electricity price and real-time electricity price data 120 days before the forecast date and the forecast power Q before adjustment 调整前 , the data are one point every 15 minutes, 96 moments in a day; Extract all data with adjustment direction of 1 within 120 days as data1, and extract all data with adjustment direction of 0 within 120 days as data2; The day-ahead clearing electricity prices in data1 and data2 are divided into buckets, and the bucket width is set to 100 yuan / MWh; Data1 and data2 are grouped and counted according to 96 moments and buckets, forming a grid with 96 moments as buckets. Each grid contains 0 or more day-ahead clearing prices, real-time clearing prices, and the predicted power Q before adjustment. 调整前 data; The day-ahead clearing price, real-time clearing price and Q 调整前 Normalization is performed to form the normalized day-ahead clearing electricity price P′ 日前 , normalized real-time clearing electricity price P′ 实时 And the normalized power prediction before regulation Q′ 调整前 ; Based on the trading rules, determine the normalized overall power forecast revenue R′ before adjustment at the current moment 调整前 and the normalized adjusted power prediction overall benefit R′ 调整后 , the specific formula is as follows: R′ 调整前 =Q′ 中长期 P′ 中长期 +(Q′ 调整前 -Q′ 中长期 )P′ 日前 +(Q′ 实时 -Q′ 调整前 )P′ 实时 R′ 调整后 =Q′ 中长期 P′ 中长期 +(Q′ 调整前 -Q′ 中长期 )P′ 日前 +(Q′ 实时 -Q′ 调整前 )P′ 实时 where Q′ 中长期 is the normalized medium and long term electricity; P′ 中长期 is the normalized medium- and long-term electricity price; Q′ 实时 is the normalized actual online power consumption; Establish the objective function f = R' 调整后 -R′ 调整前 The target result of the objective function f is used to determine the optimal adjustment ratio of the declared adjustment at the current moment. According to the benefit function, f=(Q′ 调整后 -Q′ 调整前 )*(P′ 日前 -P′ 实时 ), where Q′ 调整后 =t*Q′ 调整前 , t is the adjustment ratio, and further we can get f=(t-1)*Q′ 调整前 *(P′ 日前 -P′ 实时 ), it can be seen that the benefit is a function of the adjustment ratio; In the data set data1, the value range of the adjustment ratio t is set to [1,2.5], and the interval of the adjustment ratio is 0.
1. Specifically, for each price difference value, 16 revenue values under different adjustment ratios can be calculated. Finally, the median of all revenues corresponding to all price differences is obtained, and the adjustment ratio corresponding to the median is used as the upward adjustment ratio under the bucket at that moment. If there is no data in the bucket at that moment, the adjustment ratio is set to 1; In the data set data2, the range of the adjustment ratio t is set to [0,1), and the interval of the adjustment ratio is 0.
1. Specifically, for each price difference value, 10 revenue values under different adjustment ratios can be calculated. Finally, the median of all revenues corresponding to all price differences is obtained, and the adjustment ratio corresponding to the median is used as the downward adjustment ratio under the bucket at that moment. If there is no data in the bucket at that moment, the adjustment ratio is set to 1; This results in an upward adjustment scale grid and a downward adjustment scale grid.
9. The power curve optimization method for the spot market of a new energy station based on AI according to claim 1 is characterized in that: The method of dividing the price segments and updating the adjustment direction and adjustment probability in step S10 includes: The day-ahead predicted electricity price is divided into three categories: the lower limit of the provincial clearing price, the upper limit of the provincial clearing price, and other electricity prices that do not belong to either of the two. If the day-ahead predicted electricity price is the lower limit of the provincial clearing price, the adjustment direction is set to downward adjustment and the adjustment probability is set to 1. If the day-ahead predicted electricity price is the upper limit of the provincial clearing price, the adjustment direction is set to upward adjustment and the adjustment probability is set to 1. If the day-ahead clearing electricity price is other electricity prices, no correction is made to the adjustment direction and adjustment probability.
Citation Information
Cited By
Distributed photovoltaic power distribution network shared energy storage optimization method
CN120914866A
Multi-time-scale new energy-stored energy joint output optimization method and system based on spot price
CN121390455A
