A method for predicting day-ahead price difference in electricity spot market

Abstract

The invention relates to the technical field of price prediction in the electric power spot market, and discloses a method for predicting the day-ahead price difference in the electric power spot market, comprising: acquiring and storing historical data including electricity price data, load data, meteorological data and market dynamic data in real time; extracting key features that affect changes in electricity prices from the data; selecting multiple deep learning algorithm models as basic prediction models of an integrated model; performing cross-validation in a rolling time window manner, and dividing the historical data into a training set, a validation set and a test set; evaluating the prediction performance of the model based on the HuberMQLoss value of each basic prediction model on the validation set; and performing weighted averaging on the prediction results of multiple basic prediction models according to the prediction performance of the basic prediction model to obtain the final prediction result of the integrated day-ahead price difference prediction model, so as to solve the technical problem that the prior art is difficult to capture the nonlinear characteristics and non-steady-state changes of the day-ahead price difference.

Description

Technical Field

[0001] The present invention relates to the technical field of price prediction in the electricity spot market, and in particular to a method for predicting the day-ahead price difference in the electricity spot market. Background Art

[0002] In the power market, electricity prices are important economic signals that reflect the relationship between power supply and demand. Among them, the day-ahead electricity price and the real-time electricity price are two key types of electricity prices. The day-ahead electricity price refers to the electricity price level set by the power market operator before the start of daily trading based on the forecast of future power supply and demand. The real-time electricity price refers to the electricity price that is dynamically adjusted according to the real-time supply and demand situation when the power transaction actually occurs. The day-ahead electricity price difference is the difference between the real-time electricity price and the day-ahead electricity price; the price difference reflects the real-time supply and demand changes in the power market and the deviation between the market participants' expectations and actual situation of future power supply and demand. By calculating this price difference, we can have a deeper understanding of the dynamic operation of the power market and the sensitivity and reaction speed of market participants to changes in electricity prices.

[0003] Traditional day-ahead price difference forecasting methods mainly rely on time series analysis, such as the autoregressive moving average model (ARMA) and the autoregressive integrated moving average model (ARIMA). The basic idea of ​​these methods is to predict the trend of future electricity prices by analyzing the linear relationship in historical electricity price data. However, these methods have some obvious defects in practical applications.

[0004] These methods assume that time series data is stationary, that is, the statistical characteristics of the data (such as mean, variance, etc.) remain unchanged over time. However, the electricity spot market is a highly complex and changeable system, and its electricity price data often exhibits characteristics such as non-stationarity, nonlinearity, and high volatility. Therefore, traditional linear models are difficult to accurately capture these characteristics, resulting in inaccurate prediction results.

[0005] Secondly, existing forecasting methods are slow to respond when processing real-time market data and cannot be updated in a timely manner. Due to the extremely high real-time requirements of the power market, market participants need to be able to obtain the latest electricity price information and respond in a timely manner. However, traditional forecasting methods often require long calculation time and data update cycles, which cannot meet the real-time needs of market participants. This causes market participants to miss the best trading opportunities and cause economic losses. Summary of the invention

[0006] The present invention aims to provide a method for predicting the day-ahead price difference in the electricity spot market, so as to solve the technical problem that the prior art is difficult to capture the nonlinear characteristics and non-steady-state changes of the day-ahead price difference.

[0007] To achieve the above object, the present invention adopts the following technical solution: a method for predicting the day-ahead price difference in the electricity spot market, comprising:

[0008] Real-time acquisition and storage of historical data, including electricity price data, load data, meteorological data and market dynamic data, cleaning and preprocessing of historical data, electricity price data including historical day-ahead electricity price data and historical real-time electricity price data;

[0009] Obtain key features that affect electricity price changes from processed load data, meteorological data, and market dynamics data, including historical daily 24-hour electricity consumption, seasonal changes, and the impact of sudden weather on electricity demand, as well as the short-term impact of policy adjustments and major news events on electricity prices;

[0010] Build a day-ahead price difference prediction model, select multiple deep learning algorithm models as the basic prediction models of the integrated model, use a rolling time window to perform cross-validation, divide historical data into training set, validation set and test set, and use the training set data to train the basic prediction model, while using the validation set data to evaluate the performance of each basic prediction model;

[0011] The prediction performance of the model is evaluated based on the HuberMQLoss value of each basic prediction model on the validation set, and according to the prediction performance of the basic prediction model, the prediction results of multiple basic prediction models are weighted averaged to obtain the final prediction result of the day-ahead price difference of the integrated day-ahead price difference prediction model; the final prediction result is:

[0012]

[0013] Among them, y is the final prediction result, n is the number of basic prediction models, and y i For the prediction result of each basic prediction model, the corresponding weight is w i , w i for:

[0014]

[0015] The HuberMQLoss value of the basic prediction model on the validation set, μ is the mean of the HuberMQLoss values ​​of all basic prediction models on the validation set;

[0016] The final forecast result output is the day-ahead price difference forecast result.

[0017] The principles and advantages of this solution are: select multiple deep learning algorithm models as the basic prediction models of the integrated model. These models can capture the nonlinear characteristics and non-steady-state changes of the data. Use the rolling time window method for cross-validation, and divide the historical data into training set, validation set and test set. Use the training set data to train the basic prediction model, and use the validation set data to evaluate the performance of each basic prediction model. The prediction performance of the model is evaluated based on the HuberMQLoss value of each basic prediction model on the validation set. Then, according to the prediction performance of the basic prediction model, the prediction results of multiple basic prediction models are weighted averaged to obtain the final prediction result of the integrated day-ahead price difference prediction model. This method can combine the advantages of multiple models and improve the accuracy and stability of the prediction.

[0018] By introducing a deep learning algorithm model, this solution can capture the nonlinear characteristics and non-steady-state changes of electricity price data in the electricity spot market. The deep learning model has a strong nonlinear fitting ability and can more accurately reflect the complex changes of electricity price data. This solution not only takes into account electricity price data, but also introduces a variety of data such as load data, meteorological data, and market dynamics data. These data can reflect the supply and demand situation and market dynamics of the electricity market from different angles, providing more comprehensive information for the prediction model. By weighted averaging the prediction results of multiple basic prediction models, this solution can combine the advantages of multiple models and further improve the accuracy and stability of the prediction. At the same time, weighting according to the performance of the model on the validation set can ensure the rationality and effectiveness of the weight distribution.

[0019] Preferably, as an improvement, it also includes setting up regular updates of the day-ahead price difference prediction model to integrate new historical data; and dynamically adjusting the weights of the prediction results of each basic prediction model using an exponentially weighted average method based on the deviation between the historically predicted day-ahead price difference and the actual day-ahead price difference, specifically:

[0020] W it =α*W i(t-1) +(1-α)*E(t);

[0021] Among them, W it is the weight of the basic prediction model at time t, W i(t-1) is the weight of the basic forecast model at time t-1, α is used to control the decay rate, 0<α<1, E(t) is the actual day-ahead price difference at time t minus the forecast result of the basic forecast model at time t.

[0022] The beneficial effect of this improvement is that by dynamically adjusting the weights of each basic prediction model in the integrated model, the model can adapt to the latest data changes more quickly, thereby improving the real-time and accuracy of the prediction. Using the exponential weighted average method, the model is more sensitive to recent prediction errors and can correct prediction deviations in a timely manner to ensure the reliability of the prediction results. The method of dynamically adjusting weights can more reasonably distribute the importance of each basic prediction model in the integrated model, thereby improving the accuracy of the overall prediction.

[0023] Preferably, as an improvement, the basic prediction model includes a TimesNet model, a KAN model and an Inverted Transformer model.

[0024] The beneficial effects of this improvement are: TimesNet model: by capturing long-term dependencies and complex patterns in time series data, the model's ability to predict changes in the day-ahead price spread in the electricity market is improved; KAN model: combining the attention mechanism and external knowledge, the model's ability to understand the contextual information and potential relationships of time series data is enhanced, further improving the accuracy of the prediction; Inverted Transformer model: by reversing the connection method of the encoder and decoder, the model's ability to model long-distance dependencies in time series data is improved, thereby optimizing the prediction results.

[0025] Preferably, as an improvement, the day-ahead price difference prediction method further includes optimizing the hyperparameters of the basic prediction model using grid search and Bayesian optimization, specifically:

[0026] The HuberMQLoss loss function is defined as the objective function, and the minimization of the HuberMQLoss value is used as the goal to measure the performance of the basic prediction model under a given hyperparameter combination;

[0027] Determine the hyperparameters that need to be optimized and their value ranges; use random forest as a surrogate model to fit the functional relationship of the objective function in the hyperparameter space;

[0028] Using the confidence upper bound as the acquisition function, the next hyperparameter combination to be evaluated is selected based on the prediction results of the surrogate model;

[0029] Bayesian optimization is used for iteration. In each iteration, a new hyperparameter combination is selected for evaluation, and the proxy model and acquisition function are updated according to the evaluation results. When the maximum number of iterations is reached, the iteration process stops.

[0030] The beneficial effect of this improvement is that the grid search and Bayesian optimization methods can efficiently search for the optimal hyperparameter combination, reducing the time and cost of manual parameter adjustment. By optimizing the hyperparameters, the model can more accurately capture the characteristics of the data, thereby improving the accuracy and stability of the prediction.

[0031] Preferably, as an improvement, the hyperparameters to be optimized and their value ranges are determined, specifically:

[0032] The hyperparameters to be optimized for the TimesNet model include the number of network layers, the number of neurons in each layer, the learning rate, and the dropout rate;

[0033] The hyperparameters to be optimized for the KAN model include embedding dimension, number of attention heads, number of encoder layers, hidden layer size, learning rate, and dropout rate;

[0034] The hyperparameters to be optimized for the Inverted Transformer model include word embedding dimension, number of attention heads, number of encoder layers, size of hidden layer of feedforward network, learning rate, layer normalization parameter and dropout rate.

[0035] The beneficial effect of this improvement is that specific hyperparameters and their value ranges are determined for different basic prediction models, making the hyperparameter optimization process more clear and targeted. The clear search space reduces the number of unnecessary searches and improves the optimization efficiency.

[0036] Preferably, as an improvement, the historical data is cleaned and preprocessed, specifically including filling missing values, removing outliers and data standardization;

[0037] Fill missing values: fill missing data in historical data acquired in real time; for continuous data, use spline interpolation to fill missing values; for discrete data, use previous value filling method to fill missing values;

[0038] Eliminate outliers, draw box plots, and use statistical methods to identify data that are significantly deviated from the normal range. Data outside the upper and lower bounds of the box plot are considered outliers, which are then eliminated and the missing values ​​are filled in.

[0039] Data standardization: Use the z-score standardization method to convert historical data into a distribution with a mean of 0 and a standard deviation of 1 to eliminate the impact of dimensions between different features and standardize the data.

[0040] The beneficial effect of this improvement is that by filling missing values, removing outliers and standardizing data, the quality and consistency of historical data are improved, providing a better data foundation for model training. High-quality data can reduce noise interference in the model training process, thereby improving the accuracy and stability of predictions.

[0041] Preferably, as an improvement, the day-ahead price difference prediction method further includes presenting the final prediction result of the day-ahead price difference to market participants through a visual interface, and providing a confidence interval of the day-ahead price difference prediction.

[0042] The beneficial effects of this improvement are: the final forecast results of the day-ahead spread are presented to market participants through a visual interface, making the forecast results more intuitive and easy to understand. The confidence interval of the day-ahead spread forecast is provided, providing a more reliable basis for market participants' decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following is further described in detail through specific implementation methods:

[0045] Example

[0046] Basically as attached Figure 1 As shown, a method for predicting the day-ahead price difference in the electricity spot market includes the following steps:

[0047] Real-time acquisition of historical data, including electricity price data, load data, meteorological data and market dynamics data. API and massive data storage technology are used to ensure the real-time and accuracy of historical data.

[0048] Among them, electricity price data includes historical day-ahead electricity price data and historical real-time electricity price data, etc.; load data includes historical unified dispatching load forecasts, historical unified dispatching load actual quantities and real unified dispatching load forecasts, etc.; meteorological data includes temperature, humidity, wind speed, etc.; market dynamics data includes changes in electricity spot market policies, news, etc.

[0049] Clean and preprocess the real-time data, and obtain key features from load data, meteorological data, and market dynamics data. Data preprocessing includes: filling missing values, removing outliers, and data standardization.

[0050] Fill missing values ​​by using interpolation, mean, mode and other methods to fill missing data in real-time data. For continuous data, spline interpolation is used to fill missing values, such as meteorological data and electricity price data; for discrete data, previous value filling is used to fill missing values, such as market dynamics data.

[0051] Outliers were eliminated and box plots were drawn. Statistical methods were used to identify data that obviously deviated from the normal range. Data outside the upper and lower bounds of the box plot were considered outliers. Outliers were eliminated and missing values ​​were filled in. Outliers may be caused by data sampling errors, data transmission errors, etc.

[0052] Data standardization, using the z-score standardization method to convert the data into a distribution with a mean of 0 and a standard deviation of 1. This process eliminates the impact of dimensions between different features, so the data is standardized.

[0053] In addition to cleaning and preprocessing, word segmentation technology in the field of natural language processing is also used to extract structured data from key market quantitative information in standardized market document texts.

[0054] The key features obtained from load data, meteorological data and market dynamics data include: obtaining the historical 24-hour electricity consumption every day from load data, and identifying the peak and trough of electricity demand; obtaining the impact of seasonal changes and sudden weather on electricity demand from meteorological data; and obtaining the short-term impact of policy adjustments, major news events, etc. on electricity prices from market dynamics data.

[0055] A day-ahead price difference prediction model is constructed, and multiple deep learning algorithm models are selected as basic prediction models. In this embodiment, TimesNet, KAN and Inverted Transformer are selected as basic prediction models.

[0056] TimesNet, a deep neural network designed specifically for processing time series data, can effectively capture long-term dependencies and complex patterns in time series data.

[0057] KAN (Knowledge-Aware Attention Networks), a deep learning model that combines attention mechanism and external knowledge, can better understand the contextual information and potential relationships in time series data.

[0058] Inverted Transformer, an improved Transformer model, improves the model's ability to model long-distance dependencies of time series data by reversing the connection between the encoder and the decoder.

[0059] Cross-validation is performed using a rolling time window to divide historical data into training, validation, and test sets. The acquired historical data set is sorted in chronological order, and the data is divided into multiple subsets using a rolling time window according to the preset window size. In each window, a portion of the data is used as a training set, a portion as a validation set, and a portion as a test set.

[0060] The preset window size should be large enough to include enough data points to train the model, but small enough to reflect the changing trend of the data. The distribution principle of the training set, validation set, and test set can be set to 60%, 20%, and 20%, and the specific value can be adjusted according to the actual situation.

[0061] Use the training set data to train the basic prediction model, and use the validation set data to evaluate the performance of each basic prediction model. According to the performance of each basic prediction model on the validation set, use grid search to optimize the hyperparameters of the basic prediction model, that is, through exhaustive search in the predefined hyperparameter space, evaluate each given hyperparameter combination, and select the optimal hyperparameter combination; and use Bayesian optimization to improve the efficiency of hyperparameter optimization. The specific steps include:

[0062] Defining the objective function: First, it is necessary to define an objective function to measure the performance of the basic prediction model under a given hyperparameter combination. In the present invention, the objective function is represented by the HuberMQLoss loss function, and the goal is to minimize the HuberMQLoss value.

[0063] Select hyperparameter space: Determine the hyperparameters that need to be optimized and their value ranges. For the TimesNet model, the hyperparameters that need to be optimized include: number of network layers, number of neurons per layer, learning rate, and dropout rate; for the KAN model, the hyperparameters that need to be optimized include embedding dimension, number of attention heads, number of encoder layers, hidden layer size, learning rate, and dropout rate; for the Inverted Transformer model, the hyperparameters that need to be optimized include: word embedding dimension, number of attention heads, number of encoder layers, feedforward network hidden layer size, learning rate, layer normalization parameter, and dropout rate.

[0064] Building a surrogate model: Bayesian optimization uses a surrogate model to fit the functional relationship of the objective function in the hyperparameter space. The surrogate model used here is a random forest. The purpose of the surrogate model is to approximate the objective function with a simple function to facilitate rapid evaluation of the performance of different hyperparameter combinations.

[0065] Selecting the acquisition function: The acquisition function is used to guide the hyperparameter search process. It selects the next hyperparameter combination to evaluate based on the predictions of the surrogate model. The acquisition function used here is the upper confidence limit. The goal of the acquisition function is to strike a balance between exploration and exploitation, trying new hyperparameter combinations while making full use of existing information.

[0066] Iterative optimization: Bayesian optimization is an iterative process. Each iteration selects a new hyperparameter combination for evaluation and updates the proxy model and acquisition function based on the evaluation results. The iterative process continues until the stopping condition is met. Here, the stopping condition of reaching the maximum number of iterations is used.

[0067] The model performance is evaluated based on the HuberMQLoss value of each basic prediction model on the validation set. The lower the HuberMQLoss value, the better the prediction performance of the basic prediction model. HuberMQLoss is a loss function that combines the advantages of mean square error (MSE) and mean absolute error (MAE) and is more robust to outliers. Its expression is as follows:

[0068]

[0069] Among them, y i is the true value, f(x i ) is the predicted value, ρ δ is the Huber loss function, which is defined as follows:

[0070]

[0071] Where a = y i -f(x i ), δ is a parameter that controls the balance between MSE and MAE values ​​and is usually set to 1.

[0072] According to the prediction performance of the basic prediction model, the prediction results of multiple basic prediction models are weighted averaged to obtain the final prediction result of the day-ahead price difference of the integrated model. The final prediction result is:

[0073]

[0074] Where y is the final prediction result, n is the number of basic prediction models, in this embodiment, n=3; y i For the prediction result of each basic prediction model, the corresponding weight is w i .

[0075] Weight w i Weighted according to the validation set performance, if a model predicts well on the validation set alone, then it is given a higher weight;i for:

[0076]

[0077] The HuberMQLoss value of the basic prediction model on the validation set, μ is the mean of the HuberMQLoss values ​​of all basic prediction models on the validation set; the lower the HuberMQLoss, the higher the weight. The sum of the weights of the prediction results of all basic prediction models is 1, ensuring that the range of the final prediction result y is consistent with the prediction range of a single basic prediction model.

[0078] Update the model regularly, set the model to be updated regularly, integrate new market dynamics data and meteorological data, and in this embodiment, set the regular update time to be every natural day.

[0079] Feedback adjustment model dynamically adjusts model parameters based on the deviation between the historical predicted day-ahead spread and the actual day-ahead spread, i.e., the prediction error. The exponentially weighted average method is used to adjust the base model weights so that the model pays more attention to the recent prediction performance.

[0080] The weight of the prediction results of each basic prediction model is dynamically adjusted according to the prediction error. After each prediction, the weight of the basic prediction model in the integrated model is dynamically adjusted according to the prediction error of each basic prediction model, that is, the error value between the prediction result of the basic prediction model and the actual day-ahead price difference. The weight of the model with small error is increased, and the weight of the model with large error is reduced.

[0081] Weight adjustment method: If a model has a large error in recent predictions, we will think that its grasp of the current trend may be biased, and we need to adjust its weight more quickly to reduce its importance. In contrast, the impact of the model's prediction errors a long time ago on the present has weakened, so we will gradually reduce the weight of early errors. Here we use the exponential weighted average method. The significance of exponential weighted average:

[0082] It acts like a "filter" that helps the model gradually dilute past error information and pay more attention to recent error information. Specifically, we assign a weight to the model prediction error at each moment, and this weight decays exponentially over time. In this way, recent errors will have a larger weight, while the weight of past errors will gradually approach zero.

[0083] Assume that α (0<α<1) is used to control the decay speed. The specific value of α is set according to the requirements. it is the weight of the basic prediction model at time t, W i(t-1)is the weight of the basic forecast model at time t-1, E(t) is the error at time t, that is, the actual day-ahead price difference at time t minus the forecast result of the basic forecast model at time t. If the actual day-ahead price difference is higher than the predicted day-ahead price difference, then the forecast error E(t) is a positive number, indicating that the forecast value is too low. If the actual day-ahead price difference is lower than the predicted day-ahead price difference, then the forecast error E(t) is a negative number, indicating that the forecast value is too high. Then:

[0084] W it =α*W i(t-1) +(1-α)*E(t);

[0085] The weight at time t is not only related to the current error E(t), but also to the previous weight, and as time goes by, the impact of the previous error becomes smaller and smaller. In this way, the model can pay more attention to the recent prediction performance and adjust the weights of each base model in time, so as to better adapt to the latest data changes and improve the accuracy of the overall prediction.

[0086] The result is output, and the day-ahead price difference forecast is generated and presented to market participants through a visual interface. The confidence interval of the day-ahead price difference forecast is provided to help users assess transaction risks.

[0087] The above is only an embodiment of the present invention, and the common knowledge such as the known specific technical solutions and / or characteristics in the solution is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several modifications and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for predicting the day-ahead price difference in the electricity spot market, characterized in that: include: Real-time acquisition and storage of historical data, including electricity price data, load data, meteorological data and market dynamic data, cleaning and preprocessing of historical data, electricity price data including historical day-ahead electricity price data and historical real-time electricity price data; Obtain key features that affect electricity price changes from processed load data, meteorological data, and market dynamics data, including historical daily 24-hour electricity consumption, seasonal changes, and the impact of sudden weather on electricity demand, as well as the short-term impact of policy adjustments and major news events on electricity prices; Build a day-ahead price difference prediction model, select multiple deep learning algorithm models as the basic prediction models of the integrated model, use a rolling time window to perform cross-validation, divide historical data into training set, validation set and test set, and use the training set data to train the basic prediction model, while using the validation set data to evaluate the performance of each basic prediction model; The prediction performance of the model is evaluated based on the HuberMQLoss value of each basic prediction model on the validation set, and according to the prediction performance of the basic prediction model, the prediction results of multiple basic prediction models are weighted averaged to obtain the final prediction result of the day-ahead price difference of the integrated day-ahead price difference prediction model; the final prediction result is: Among them, y is the final prediction result, n is the number of basic prediction models, and y i For the prediction result of each basic prediction model, the corresponding weight is w i , w i for: The HuberMQLoss value of the basic prediction model on the validation set, μ is the mean of the HuberMQLoss values ​​of all basic prediction models on the validation set; The final forecast result output is the day-ahead price difference forecast result.

2. A method for predicting the day-ahead price difference in the electricity spot market according to claim 1, characterized in that: It also includes setting up regular updates of the day-ahead spread forecasting model and integrating new historical data; and dynamically adjusting the weights of the forecast results of each basic forecasting model using the exponential weighted average method based on the deviation between the historical forecast day-ahead spread and the actual day-ahead spread, specifically: W it =α*W i(t-1) +(1-α)*E(t); Among them, W it is the weight of the basic prediction model at time t, W i(t-1) is the weight of the basic forecast model at time t-1, α is used to control the decay rate, 0<α<1, E(t) is the actual day-ahead price difference at time t minus the forecast result of the basic forecast model at time t.

3. The method for predicting the day-ahead price difference in the electricity spot market according to claim 2, characterized in that: The basic prediction models include TimesNet model, KAN model and Inverted Transformer model.

4. The method for predicting the day-ahead price difference in the electricity spot market according to claim 3, characterized in that: The day-ahead price difference prediction method further includes optimizing the hyperparameters of the basic prediction model using grid search and Bayesian optimization, specifically: The HuberMQLoss loss function is defined as the objective function, and the minimization of the HuberMQLoss value is used as the goal to measure the performance of the basic prediction model under a given hyperparameter combination; Determine the hyperparameters that need to be optimized and their value ranges; use random forest as a surrogate model to fit the functional relationship of the objective function in the hyperparameter space; Using the confidence upper bound as the acquisition function, the next hyperparameter combination to be evaluated is selected based on the prediction results of the surrogate model; Bayesian optimization is used for iteration. In each iteration, a new hyperparameter combination is selected for evaluation, and the proxy model and acquisition function are updated according to the evaluation results. When the maximum number of iterations is reached, the iteration process stops.

5. A method for predicting the day-ahead price difference in the electricity spot market according to claim 4, characterized in that: Determine the hyperparameters that need to be optimized and their value ranges, specifically: The hyperparameters to be optimized for the TimesNet model include the number of network layers, the number of neurons in each layer, the learning rate, and the dropout rate; The hyperparameters to be optimized for the KAN model include embedding dimension, number of attention heads, number of encoder layers, hidden layer size, learning rate, and dropout rate; The hyperparameters to be optimized for the Inverted Transformer model include word embedding dimension, number of attention heads, number of encoder layers, size of hidden layer of feedforward network, learning rate, layer normalization parameter and dropout rate.

6. A method for predicting the day-ahead price difference in the electricity spot market according to claim 5, characterized in that: Clean and preprocess historical data, including filling missing values, removing outliers, and standardizing data; Fill missing values: fill missing data in historical data acquired in real time; for continuous data, use spline interpolation to fill missing values; for discrete data, use previous value filling method to fill missing values; Eliminate outliers, draw box plots, and use statistical methods to identify data that are significantly deviated from the normal range. Data outside the upper and lower bounds of the box plot are considered outliers, which are then eliminated and the missing values ​​are filled in. Data standardization: Use the z-score standardization method to convert historical data into a distribution with a mean of 0 and a standard deviation of 1 to eliminate the impact of dimensions between different features and standardize the data.

7. A method for predicting day-ahead price difference in the electricity spot market according to claim 6, characterized in that: The day-ahead price difference prediction method also includes presenting the final prediction result of the day-ahead price difference to market participants through a visual interface, and providing a confidence interval of the day-ahead price difference prediction.

Images