Coal price prediction system based on multi-feature fusion LSTM network

Abstract

The invention relates to the technical field of coal price prediction, in particular to a coal price prediction system based on a multi-feature fusion LSTM network, which comprises a central prediction module and at least one local data acquisition module in communication connection with the central prediction module, any local data acquisition module is used for acquiring a to-be-predicted coal price data set in a first preset area and sending the to-be-predicted coal price data set to the central prediction module; the central prediction module is used for inputting a coal price data set into a preset coal price prediction model when receiving the coal price data set sent by any local data acquisition module, obtaining a corresponding coal price prediction result, and sending the coal price prediction result to the corresponding local data acquisition module. And the local data acquisition module is assisted to make business decisions. The method has the beneficial effect that the dominant position and risk response capability of an enterprise in market competition are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of coal price prediction, and in particular to a coal price prediction system based on a multi-feature fusion LSTM network. Background Art

[0002] In the field of traditional coal market price forecasting, time series analysis, conventional regression models and forecasting methods driven by traditional economic theories have long been dominant, but they have obvious limitations in the modern market environment.

[0003] Traditional methods rely on a relatively simple data foundation, relying on historical coal price data for modeling and ignoring the complex and multifaceted influences of the coal price formation mechanism. However, coal price fluctuations are deeply intertwined and synergistically influenced by factors such as market demand dynamics (such as industrial production activity cycles, seasonal differences in power sector electricity consumption, and changes in residential heating demand), supply chain changes (coal mining and transportation efficiency, port inventory and logistics efficiency, etc.), macroeconomic conditions (GDP growth, monetary policy, industrial value added, exchange rates, etc.), and weather conditions (such as sudden changes in energy demand caused by extreme weather and the impact of natural disasters on coal production, sales, and transportation).

[0004] Due to the neglect of the above factors, traditional methods cannot accurately reflect the inherent complexity and dynamics of the coal price formation mechanism, which restricts the accuracy and reliability of the forecast and makes it difficult to meet the urgent needs of market participants for accurate price forecasts.

[0005] Furthermore, coal price formation follows a complex nonlinear relationship, with various influencing factors intertwined to form an intricate nonlinear correlation. Traditional forecasting methods, based on linear assumptions or simple functional relationships, struggle to accurately capture the essential characteristics and dynamic changes of these nonlinear relationships. Consequently, they are unable to accurately track coal price trends in a complex and volatile market environment, resulting in significant forecasting errors. These methods are gradually losing their value given the demand for high-precision forecasts in the modern coal market.

[0006] Moreover, the coal market is changing rapidly. Traditional forecasting methods are limited by their own model architecture and algorithms, and have lags in data updates and model adjustments. They are unable to capture market dynamics and feed them back into forecast results in a timely manner, making it difficult to provide market participants with timely and effective decision-making basis in the rapidly changing market rhythm. Summary of the Invention

[0007] (1) Technical issues to be resolved

[0008] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a coal price forecasting system based on a multi-feature fusion LSTM network, which solves the technical problems that traditional forecasting methods have lags in data updating and model adjustment, are unable to capture market dynamic changes in a timely manner and feed them back to the forecast results, and are difficult to provide market participants with timely and effective decision-making basis in the rapidly changing market rhythm.

[0009] (2) Technical solution

[0010] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:

[0011] In a first aspect, an embodiment of the present invention provides a coal price prediction system based on a multi-feature fusion LSTM network, comprising: a central prediction module, and at least one local data acquisition module in communication with the central prediction module;

[0012] Any of the local data acquisition modules is used to obtain a data set of coal prices to be predicted in a first preset area, and send the data set of coal prices to be predicted to the central prediction module;

[0013] The central prediction module is configured to, upon receiving a coal price data set sent by any local data collection module, input the coal price data set into a pre-set coal price prediction model to obtain a corresponding coal price prediction result, and send the coal price prediction result to the corresponding local data collection module to assist the local data collection module in making business decisions;

[0014] The coal price prediction model includes: a prediction model obtained by training an LSTM network through a prediction training set;

[0015] The prediction training set is a training set obtained by performing feature screening on the original data of the second preset area acquired in advance according to a preset feature screening algorithm, and based on the original data after feature screening and a preset first hysteresis strategy;

[0016] The original data includes the coal price corresponding to each unit time in the first preset time period and the coal price influencing characteristics affecting the change of the coal price;

[0017] The first hysteresis strategy is to establish a mapping relationship between the coal price corresponding to any unit time in the original data and the historical coal price data corresponding to a second preset time past the unit time.

[0018] Optionally, the central prediction module includes: a model training unit and a central prediction unit;

[0019] The model training unit is used to pre-process the pre-acquired raw data of the second preset area;

[0020] Performing feature screening on the pre-acquired raw data according to a pre-set feature screening algorithm to reduce the feature dimension of the coal price impact feature corresponding to the coal price in the raw data; the coal price impact feature includes at least one feature dimension;

[0021] Based on the original data after feature screening and the pre-set first lag strategy, a corresponding prediction training set is obtained;

[0022] Training a pre-built LSTM network based on the prediction training set to obtain a trained coal price prediction model;

[0023] The central prediction unit is used to input the coal price data group into a preset coal price prediction model when receiving the coal price data group sent by any local data acquisition module, obtain the corresponding coal price prediction result, and send the coal price prediction result to the corresponding local data acquisition module to assist the local data acquisition module in making business decisions.

[0024] Optionally, the model training unit pre-processes the pre-acquired raw data of the second preset area, including:

[0025] performing abnormal data detection on the pre-acquired raw data of the second preset area according to a preset Z-score algorithm, marking abnormal data in the raw data, and replacing the marked abnormal data according to a preset abnormal data replacement algorithm;

[0026] The abnormal data replacement algorithm includes data difference and / or median replacement.

[0027] Optionally, the model training unit performs feature screening on the pre-acquired raw data according to a pre-set feature screening algorithm, including:

[0028] According to the preset Pearson correlation analysis algorithm, Lasso regression algorithm, grey relational analysis algorithm and random forest regression algorithm, feature screening is performed on the coal price impact features corresponding to the coal price in the original data;

[0029] The coal price impact characteristics retained after filtering the original data using Pearson correlation analysis, Lasso regression, grey correlation analysis, and random forest regression were obtained to reduce the characteristic dimensions of the coal price impact characteristics corresponding to the coal prices in the original data.

[0030] Optionally, the model training unit performs feature screening on coal price impact features corresponding to the coal prices in the original data according to a preset Pearson correlation analysis algorithm, a Lasso regression algorithm, a grey correlation analysis algorithm, and a random forest regression algorithm, including:

[0031] Obtaining the correlation coefficient between the coal price influencing feature corresponding to each feature dimension and the coal price based on the original data and a preset Pearson correlation analysis algorithm;

[0032] Based on the correlation coefficient between the coal price impact feature corresponding to each feature dimension and the coal price and a preset first correlation coefficient threshold, feature screening is performed on the coal price impact feature corresponding to the coal price in the original data: the Pearson correlation analysis includes:

[0033]

[0034] Among them, x i is the value of the coal price impact feature of any feature dimension at the i-th unit time, y i is the coal price corresponding to the i-th unit time, is the average value of the coal price impact feature of any characteristic dimension in n unit time, is the average coal price in n unit time, r is the correlation coefficient between the coal price influencing feature of this feature dimension and the coal price;

[0035] Based on the original data and the preset Lasso regression algorithm, feature screening is performed on the coal price impact features corresponding to the coal price in the original data; the objective function corresponding to the Lasso regression algorithm is:

[0036]

[0037] Where y is the corresponding coal price matrix in the original data, X is the feature matrix corresponding to the coal price impact feature of any feature dimension, β is the preset regression coefficient, and α is the preset regularization parameter;

[0038] Obtaining, based on the original data, a feature sequence corresponding to the coal price impact feature of each feature dimension and a coal price sequence in the original data;

[0039] Obtaining the correlation between the coal price influencing feature corresponding to each feature dimension and the coal price based on the feature sequence, the coal price sequence and a preset grey correlation analysis algorithm;

[0040] According to the correlation between the coal price influencing feature corresponding to each feature dimension and the coal price and a preset second correlation threshold, the coal price influencing feature corresponding to the coal price in the original data is subjected to feature screening: the grey correlation analysis algorithm includes:

[0041]

[0042] Where c is the correlation between the coal price impact feature corresponding to any feature dimension and the coal price, Δ(x o ,x o) is the difference between the characteristic series and the coal price series;

[0043] According to a preset random forest regression algorithm, feature screening is performed on coal price impact features corresponding to the coal prices in the original data.

[0044] Optionally, the model training unit obtains a corresponding prediction training set based on the original data after feature screening and a preset first hysteresis strategy, including:

[0045] Based on a preset first hysteresis strategy, a mapping relationship is established between the coal price corresponding to any unit time in the original data and the historical coal price data corresponding to a second preset time past the unit time;

[0046] Obtaining a mapping data set based on a mapping relationship between the coal price corresponding to each unit time in the original data and the historical coal price data corresponding to a second preset time past the unit time;

[0047] The mapping dataset is divided into a prediction training set and a prediction test set in a ratio of 8:2.

[0048] Optionally, the model training unit trains a pre-built LSTM network based on the prediction training set to obtain a trained coal price prediction model, including:

[0049] The pre-built LSTM network is trained based on the prediction training set. During the training process, the corresponding loss function is obtained in real time according to the pre-set Adam optimizer and mean square error algorithm to adjust the hyperparameters of the coal price prediction model in real time to obtain a trained coal price prediction model;

[0050] Then, the model training unit is further used to:

[0051] The coal price prediction model is tested based on a pre-set prediction test set to determine whether it is qualified.

[0052] Optionally, when the central prediction module receives a coal price data set sent by any local data acquisition module, it inputs the coal price data set into a preset coal price prediction model to obtain a corresponding coal price prediction result, including:

[0053] When a coal price data group is received from any local data acquisition module, a corresponding time coding feature is assigned to the coal price of each unit time in the coal price data group according to the timestamp corresponding to each unit time in the received coal price data group and a preset formula 1; the coal price data group includes the coal price and timestamp corresponding to each unit time in the third preset time period; the formula 1 is:

[0054]

[0055] Wherein, sin(Z) and cos(Z) are the time coding features corresponding to the coal price of any unit time in the coal price data set, Z is the preset time period, and l is the period constant matched with the time period;

[0056] According to the pre-set data extrapolation strategy, the extrapolation characteristics corresponding to the coal price per unit time in the coal price data group are obtained;

[0057] Dividing the coal price data group into at least one time window according to a time series according to a preset sliding window;

[0058] According to the pre-set attention mechanism, corresponding attention weights are assigned to the coal prices of all unit times in each time window;

[0059] According to the coal price of each unit time in all time windows, as well as the time coding features, extrapolation features and attention weights corresponding to the coal price, the prediction results of the market coal price corresponding to the coal price data group are obtained.

[0060] Optionally, the data extrapolation strategy includes a second lag sub-strategy, a rolling statistics sub-strategy and a differential sub-strategy;

[0061] Then, the central prediction module obtains the extrapolation features corresponding to the coal price per unit time in the coal price data set according to the pre-set data extrapolation strategy, including:

[0062] According to the coal price data set and a preset second hysteresis sub-strategy, obtaining a hysteresis feature corresponding to the coal price per unit time in the coal price data set;

[0063] The hysteresis feature is the coal price corresponding to all unit times within the fourth preset period past the unit time corresponding to the coal price at any unit time in the coal price data group;

[0064] According to the preset rolling statistical sub-strategy, the rolling characteristics corresponding to the coal price per unit time in the coal price data group are obtained;

[0065] The rolling characteristics include the coal price of any unit time in the coal price data set, and the mean, standard deviation, maximum value and minimum value of the coal prices corresponding to all unit times in the fifth preset period past the corresponding unit time;

[0066] According to the pre-set differential sub-strategy, the differential features corresponding to the coal price of each unit time in the coal price data set are obtained;

[0067] The differential feature includes the difference between the coal price at any unit time in the coal price data set and the coal price corresponding to the unit time that is six preset times past the unit time.

[0068] Optionally, any of the local data acquisition modules is set in a corresponding coal-fired power plant; the first preset area is the coal-fired power plant where the local data acquisition module is set;

[0069] Any of the local data acquisition modules is further configured to:

[0070] When the coal price forecast result sent by the central forecast module is received, the coal-fired electricity cost in the time period corresponding to the coal price forecast result is obtained according to the coal price forecast result.

[0071] (3) Beneficial effects

[0072] The beneficial effect of the present invention is that the coal price prediction system based on the multi-feature fusion LSTM network provided by the present invention, due to the integration of the coal price prediction model of multiple feature dimensions, is more effective in assisting in formulating scientific, reasonable and forward-looking decision-making strategies in a complex and changing market environment compared with the existing technology, thereby significantly improving the company's dominant position and risk response capabilities in market competition. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 A schematic diagram of the structure of a coal price prediction system based on a multi-feature fusion LSTM network provided by an embodiment of the present invention;

[0074] Figure 2 A flow chart of a coal price prediction model training method for a multi-feature fusion LSTM network provided in an embodiment of the present invention;

[0075] Figure 3 A schematic diagram of the Pearson correlation analysis algorithm provided by an embodiment of the present invention;

[0076] Figure 4 Schematic diagram of Lasso regression algorithm analysis provided by an embodiment of the present invention;

[0077] Figure 5 Schematic diagram of the random forest regression algorithm analysis provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0078] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.

[0079] The coal price prediction system based on the multi-feature fusion LSTM network proposed in the embodiment of the present invention integrates a coal price prediction model with multiple feature dimensions. Compared with the existing technology, it effectively assists in formulating scientific, reasonable and forward-looking decision-making strategies in a complex and changing market environment, thereby significantly improving the company's dominant position and risk response capabilities in market competition.

[0080] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. Instead, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0081] Example 1

[0082] The embodiment of the present invention provides a coal price prediction system based on a multi-feature fusion LSTM network. Figure 1 As shown, it includes a central prediction module and at least one local data acquisition module in communication with the central prediction module;

[0083] Any of the local data acquisition modules is used to obtain a data set of coal prices to be predicted in a first preset area, and send the data set of coal prices to be predicted to the central prediction module;

[0084] The central prediction module is configured to, upon receiving a coal price data set sent by any local data collection module, input the coal price data set into a pre-set coal price prediction model to obtain a corresponding coal price prediction result, and send the coal price prediction result to the corresponding local data collection module to assist the local data collection module in making business decisions;

[0085] The coal price prediction model includes: a prediction model obtained by training an LSTM network through a prediction training set;

[0086] The prediction training set is a training set obtained by performing feature screening on the original data of the second preset area acquired in advance according to a preset feature screening algorithm, and based on the original data after feature screening and a preset first hysteresis strategy;

[0087] The original data includes the coal price corresponding to each unit time in the first preset time period and the coal price influencing characteristics affecting the change of the coal price;

[0088] The first hysteresis strategy is to establish a mapping relationship between the coal price corresponding to any unit time in the original data and the historical coal price data corresponding to a second preset time past the unit time.

[0089] The coal price prediction system based on the multi-feature fusion LSTM network proposed in the embodiment of the present invention integrates a coal price prediction model with multiple feature dimensions. Compared with the existing technology, it effectively assists in formulating scientific, reasonable and forward-looking decision-making strategies in a complex and changing market environment, thereby significantly improving the company's dominant position and risk response capabilities in market competition.

[0090] Example 2

[0091] This embodiment provides a coal price prediction model training method for a multi-feature fusion LSTM network. Figure 2 Shown, including:

[0092] S1, data collection and preprocessing;

[0093] Data Collection: Raw data is collected based on a pre-set data platform. The data platform connects with authoritative databases such as the National Bureau of Statistics, the China Coal Transportation and Sales Association, and the International Energy Agency to obtain relevant statistical data, which is then filtered to form raw data. For example, data on thermal coal prices, macroeconomics, inventory, alternative resources, international coal prices, transportation, and consumption from January 1, 2015, to December 16, 2024 are collected through the data platform to generate corresponding raw data. The raw data corresponding to January 1, 2015, to December 16, 2024 are shown in Table 1:

[0094] Table 1: Raw data from January 1, 2015 to December 16, 2024

[0095]

[0096]

[0097] Data preprocessing: Outliers are values in a dataset that significantly exceed the range of the rest of the data. These values may be caused by manual input errors, measurement errors, or unusual events. If neglected, they can lead to erroneous results in certain modeling scenarios. Therefore, during data cleaning, outliers must be identified and addressed to ensure model accuracy.

[0098] The standardized Z-score algorithm is used to detect outliers in the original data and calculate the Z-score value of each data point. The formula corresponding to the Z-score algorithm is as follows:

[0099]

[0100] Where x is the value of the coal price or the coal price influencing feature of any characteristic dimension in the original data at any unit time, μ is the data mean of the coal price or the coal price influencing feature of any characteristic dimension, and σ is the data standard deviation of the coal price or the coal price influencing feature of any characteristic dimension;

[0101] When |Z|>3, the value of the unit time is judged to be an abnormal value and is processed using algorithms such as data difference or median replacement.

[0102] S2. Feature screening: By comprehensively using algorithms such as Pearson correlation analysis algorithm, Lasso regression algorithm, grey correlation analysis algorithm and random forest regression algorithm, we can accurately screen out key features closely related to coal prices from the original data and take the intersection to effectively reduce the data dimension and improve the model efficiency.

[0103] The Pearson correlation analysis algorithm is a commonly used statistical algorithm for measuring the linear relationship between two variables. In coal price prediction models, using the Pearson correlation analysis algorithm can help identify the linear relationship between features (features influencing coal prices) and the target variable (i.e., coal prices). By calculating the correlation coefficient between each feature and the coal price, we can identify those with a strong correlation with coal price fluctuations, thereby improving the model's prediction accuracy.

[0104] The specific processing process is as follows: Figure 3 As shown: according to the original data and the preset Pearson correlation analysis algorithm, the correlation coefficient between the coal price influencing feature corresponding to each feature dimension and the coal price is obtained;

[0105] Based on the correlation coefficient between the coal price impact feature corresponding to each feature dimension and the coal price and a preset first correlation coefficient threshold, feature screening is performed on the coal price impact feature corresponding to the coal price in the original data: the Pearson correlation analysis includes:

[0106]

[0107] Among them, x i is the value of the coal price impact feature of any feature dimension at the i-th unit time, y i is the coal price corresponding to the i-th unit time, is the average value of the coal price impact feature of any characteristic dimension in n unit time, is the average coal price in n unit time, and r is the correlation coefficient between the coal price impact feature of this feature dimension and the coal price.

[0108] The Lasso regression algorithm (Least Absolute Shrinkage and Selection Operator) is a regression algorithm commonly used for feature selection and regularization. Compared with traditional linear regression, the Lasso regression algorithm introduces the L1 regularization term to constrain the model. It has the following characteristics and advantages and is particularly suitable for problems such as coal price prediction: (1) Feature selection: An important feature of Lasso regression is that it can automatically perform feature selection. In the regression model, the L1 regularization term penalizes the absolute value of the regression coefficient, which can shrink the regression coefficients of some unimportant features to zero, thereby achieving the purpose of eliminating these features. For the coal price prediction model, if some features do not significantly affect the coal price, Lasso regression will automatically adjust the weights of these features to zero, thereby simplifying the model and reducing overfitting.

[0109] Improve model generalization: By constraining model parameters, the Lasso regression algorithm reduces complexity, thereby improving the model's generalization and avoiding overfitting. Especially when there are a large number of features, the Lasso regression algorithm can effectively improve the model's performance on the test set.

[0110] Applicable to high-dimensional data: When the feature dimension is high, the Lasso regression algorithm can still effectively select important features through regularization, reduce unnecessary calculations and redundant features, and is particularly suitable for multi-dimensional data sets such as coal price prediction.

[0111] Improve prediction accuracy: Since the Lasso regression algorithm can effectively remove redundant features, it can usually improve prediction accuracy in some tasks, especially when there is multicollinearity between features. The Lasso regression algorithm can reduce the bias of the model.

[0112] The specific process is as follows: Figure 4 As shown: Based on the original data and the preset Lasso regression algorithm, the coal price impact features corresponding to the coal price in the original data are screened; the objective function corresponding to the Lasso regression algorithm is:

[0113]

[0114] Where y is the coal price matrix corresponding to the original data, X is the feature matrix corresponding to the coal price influencing feature of any feature dimension, β is the pre-set regression coefficient, α is the pre-set regularization parameter, and the optimal α value is 1.0;

[0115] Grey Relational Analysis (GRA) is a statistical algorithm used to reveal the relationships between multiple variables. It is often used for quantitative analysis of relationships between data, and is particularly suitable for situations where the system is highly complex and the data is incomplete or uncertain. The core idea of the grey relational analysis algorithm is to reveal the relationship between variables by comparing the correlation between each factor and the target. The advantages of using the grey relational analysis algorithm in the coal price forecasting model include:

[0116] Handling uncertainty and incomplete data: Grey relational analysis algorithms do not require complete data or a normal distribution as required by traditional statistical methods. They can effectively handle incomplete or uncertain data. In coal price forecasts, there may be some missing data, noise, or uncertainty. Grey relational analysis algorithms can help extract valid information in these situations.

[0117] Uncovering relationships between variables: The grey correlation analysis algorithm identifies the factors most closely related to coal prices by calculating the correlation between each feature (features influencing coal prices) and the target variable (e.g., coal prices). It reveals the dynamic relationships between variables and helps select the features most closely associated with coal price fluctuations.

[0118] Avoid complex assumptions and models: Unlike other algorithms (such as Pearson correlation or regression analysis), the grey relational analysis algorithm does not require the data to meet specific statistical assumptions (such as linear relationship, normal distribution, etc.). It is suitable for data that do not meet these assumptions, especially for data sets with high system complexity and nonlinear relationships between variables.

[0119] Efficient calculation and intuitiveness: The calculation of the grey relational analysis algorithm is relatively simple. It can quickly calculate the correlation between each feature and the target variable and sort the features according to the size of the correlation.

[0120] The specific processing process is: obtaining the feature sequence corresponding to the coal price impact feature of each feature dimension and the coal price sequence in the original data according to the original data;

[0121] Obtaining the correlation between the coal price influencing feature corresponding to each feature dimension and the coal price based on the feature sequence, the coal price sequence and a preset grey correlation analysis algorithm;

[0122] According to the correlation between the coal price influencing feature corresponding to each feature dimension and the coal price and a preset second correlation threshold, the coal price influencing feature corresponding to the coal price in the original data is subjected to feature screening: the grey correlation analysis algorithm includes:

[0123]

[0124] Where c is the correlation between the coal price impact feature corresponding to any feature dimension and the coal price, Δ(x o ,x o ) is the difference between the characteristic sequence and the coal price sequence.

[0125] Random Forest Regression is an ensemble learning algorithm that makes predictions by building multiple decision trees and averaging their outputs. Random Forest Regression has powerful predictive capabilities, particularly for high-dimensional data and complex nonlinear relationships. Using Random Forest Regression for tasks like coal price forecasting offers several significant advantages:

[0126] Handling nonlinear relationships:

[0127] The random forest regression algorithm is a nonlinear model that effectively captures the complex nonlinear relationships between features and target variables. In coal price forecasting, the relationship between features and price may not be a simple linear one. Random forests can automatically model and learn these complex relationships, thereby improving forecast accuracy.

[0128] No feature preprocessing required: The random forest regression algorithm does not require standardized or normalized features and can handle features of different scales. For various features in the coal price prediction model (such as coal type, market demand, transportation cost, etc.), the random forest regression algorithm can directly use the raw data for training without complex preprocessing.

[0129] Strong resistance to overfitting: As the number of decision trees increases, the random forest regression algorithm effectively avoids the overfitting problem of a single decision tree. By introducing randomness (for example, randomly selecting features and samples), the random forest ensures that each tree fits the data independently. Ultimately, the predictions of these decision trees are combined through majority voting or averaging, enhancing the model's generalization ability.

[0130] Important Feature Assessment: Random forest regression can assess the importance of each feature, which is particularly useful for feature selection in coal price forecasting. By calculating feature importance scores, we can identify which features have the greatest impact on coal price changes, providing a basis for feature selection and model optimization.

[0131] The specific process is as follows: Figure 5 As shown, according to the preset random forest regression algorithm, the coal price impact features corresponding to the coal prices in the original data are feature screened.

[0132] Subsequently, the coal price impact features retained after filtering the original data using Pearson correlation analysis, Lasso regression, grey correlation analysis, and random forest regression were obtained to reduce the feature dimension of the coal price impact features corresponding to the coal prices in the original data. The original data after feature filtering is shown in Table 2:

[0133] Table 2: Raw data after feature screening

[0134]

[0135]

[0136] S3. Model construction and training: Build a prediction model based on a long short-term memory (LSTM) network. Create hysteresis features through a "lag" operation. Split the dataset into 80% training and 20% test sets. The model includes LSTM layers and fully connected layers with specific parameters. Compile and train using the Adam optimizer and mean squared error loss function to enable the model to deeply learn data patterns and optimize parameters. The specific steps are as follows:

[0137] Because coal prices represent time series data, training models like LSTMs require historical data to predict future values. Therefore, a "lag" operation is required to generate historical feature data. The create_lag_features function implements this process. It lags the target variable (coal price) and feature data, generating data from multiple historical moments as input features for the current moment. Lagged target variables: For example, coal prices lagged seven days can be used as a basis for predicting the current coal price. Lagged feature variables: Correspondingly, feature variables influencing coal prices are also processed based on the number of lags and serve as input features for the model.

[0138] Split the dataset into a training set and a test set. Using 80% of the data for training and 20% for testing is a common practice in machine learning. This partitioning ensures that the model learns sufficient information during training while also allowing validation on the test set to verify its generalization capabilities.

[0139] Construction of the LSTM model. LSTM (Long Short-Term Memory) is a deep learning model specifically designed for processing time series data. By introducing a "gating" mechanism, it can remember long-term dependencies and avoid the vanishing gradient problem that occurs in traditional RNNs. LSTM layer: The first layer of the model is an LSTM layer containing 50 neurons. activation = 'relu' indicates the use of the ReLU activation function. The input shape of the LSTM layer is (time steps, features), meaning that each sample has 7 days of historical data and corresponding features. Dense layer: A fully connected layer (Dense layer) is added after the LSTM layer, outputting a single predicted value, namely the future coal price. Compiling the model: Using the Adam optimizer and mean squared error (MSE) as the loss function. The Adam optimizer is an optimization algorithm suitable for deep learning and has good convergence properties.

[0140] Model training: `model_lstm.fit` is used to train the model. `epochs=100` indicates that the model will be trained for 100 cycles, and `batch_size=32` specifies that 32 samples are used for each gradient update. During training, the LSTM model adjusts its internal parameters (such as weights and biases) to minimize prediction error. After training, `model_lstm.predict` is used to predict the model on the test set. Since the test set data was not used in training, the prediction results can be used to evaluate the model's generalization ability.

[0141] Evaluation is an important step in machine learning. Its purpose is to measure the performance of the model through some indicators. The model is compiled using the Adam optimizer and the mean squared error (MSE) as the loss function.

[0142]

[0143] Here, θ is a parameter of the model. This formula indicates that the goal of model training is to find the value of parameter θ that minimizes the mean squared error (MSE). Using the Adam optimizer and the mean squared error (MSE) as the loss function, the model performs gradient updates on the training set.

[0144] Finally, the trained coal price prediction model is obtained.

[0145] S4, extrapolation model training;

[0146] Further optimize the model performance, extract time features such as months, weeks, and quarters and perform nonlinear expressions using sine and cosine functions, add lag features, rolling statistical features, and differential features, use sliding window technology to transform data, and introduce an attention mechanism so that the model can automatically assign weights according to the importance of different time steps, thereby improving the modeling and prediction capabilities of complex time series.

[0147] The specific steps are as follows:

[0148] Extracting time features: By extracting time information such as month, week, and quarter, the model can capture these periodic features. At the same time, to enhance the model's nonlinear expression capabilities, sine (sin) and cosine (cos) are used to represent periodic time features. Extract time information such as month M, week W, and quarter Q, and encode them using sine and cosine functions:

[0149] M is month, W is week, Q is quarter

[0150]

[0151] Lag features are commonly used in time series modeling. They use past values of the target variable as input features for the current time. For example, the price of the past seven days influences today's price. By incorporating past values, the model can capture historical information.

[0152] Moving Statistics: Moving Statistics calculates statistics of the target variable within a fixed time window to reflect data trends and fluctuations within that time period. These statistics, such as mean (MA), maximum, minimum, and standard deviation, summarize the changing patterns of time series data, helping the model identify periods of greater volatility and thus making more accurate forecasts.

[0153]

[0154] Max k =max(P t-k+1 ,…,P t )

[0155] Min k =min(P t-k+1 ,…,P t )

[0156] Among them, P i is the coal price on the i-th day, and k is the size of the rolling window.

[0157] Differencing is a common method in time series analysis. It captures the changing patterns of a time series by calculating the difference between the current time point and the previous time point (or a lagged time point). Differencing can help models eliminate trends and seasonality, making time series more stable and enabling better modeling and forecasting.

[0158] Diff k =P t -P t-k ;

[0159] Among them, P t is the coal price at the current time point, P t-k is the coal price at the previous k time points.

[0160] Sliding window segmentation is a common technique in time series forecasting, used to convert time series data into input-output pairs suitable for model training and prediction. Its main function is to provide the model with time-dependent features by dividing the time series data into multiple subsequences (windows) of fixed length. Sliding windows can help us capture time series.

[0161] Attention Mechanism: In time series forecasting, data at different time steps contribute differently to the current forecast. Some time steps may contain more critical information and have a greater impact on the forecast, while other time steps may only provide a relatively small contribution. Traditional time series models (such as RNN, LSTM, etc.) usually process the input of all time steps in a fixed manner and cannot clearly distinguish which time steps are more important for the forecast. The introduction of the attention mechanism aims to solve this problem, allowing the model to automatically learn to focus on those time points that have a greater impact on the current task when processing each time step.

[0162] Focusing on Important Timesteps: The attention mechanism assigns different "weights" to each timestep, meaning the model automatically determines the importance of each timestep based on the context of the current input. This allows the model to adaptively adjust its focus on historical data based on different prediction tasks, thereby improving prediction accuracy. For example, in coal price forecasting, price fluctuations in recent days may be particularly important for predicting future prices, while price changes at other times may have less impact. Adaptive Weighting: The attention mechanism determines which inputs the model should focus on by calculating an "attention weight" for each timestep. Specifically, attention weights are calculated using a similarity metric (such as dot-product attention) that reflects the relationship between the current timestep and historical timesteps. Timesteps with higher weights are given more attention, while those with lower weights are ignored. Enhanced Model Performance: The introduction of the attention mechanism not only allows the model to more accurately capture timesteps that are important for prediction, but also effectively reduces noise interference, improving its ability to model complex temporal dependencies. Furthermore, the attention mechanism enables the model to avoid long-distance dependencies when processing long time series.

[0163] The coal price prediction results are obtained through the trained coal price prediction model, and the coal price prediction results are shown in Table 3:

[0164] Table 3: Coal price forecast results

[0165]

[0166]

[0167] This embodiment provides a coal price prediction model training method for a multi-feature fusion LSTM network. It comprehensively utilizes multiple feature screening methods to effectively remove redundant information, improve model operation efficiency, and reduce the consumption of computing resources. Combined with rich time series feature engineering techniques, it can more accurately capture the long-term dependencies and complex change patterns of coal prices. The introduction of the attention mechanism enables the model to adaptively focus on important information, further enhancing the ability to model complex time series and providing highly reliable coal price forecast results for energy market participants. The model can respond to market changes in real time and adapt to complex and changing market environments. It helps enterprises to rationally plan production, procurement, and sales strategies, reduce operating risks, and improve economic efficiency. It plays an important role in promoting the stable operation and sustainable development of the energy market.

[0168] Example 3

[0169] This embodiment provides a coal price prediction system using a multi-feature fused LSTM network. The system is implemented based on the coal price prediction model trained in Example 2.

[0170] The system includes: a central prediction module, and at least one local data acquisition module in communication with the central prediction module; each local data acquisition module is deployed in a corresponding coal-fired power plant;

[0171] Any of the local data collection modules is used to collect the coal price data group to be predicted for the corresponding coal-fired power plant, and send the coal price data group to be predicted to the central prediction module;

[0172] The central prediction module is configured to, upon receiving a coal price data set sent by any local data collection module, input the coal price data set into a pre-set coal price prediction model to obtain a corresponding coal price prediction result, and send the coal price prediction result to the corresponding local data collection module to assist the local data collection module in making business decisions;

[0173] Furthermore, any of the local data collection modules is further configured to: upon receiving a coal price forecast result sent by the central forecast module, obtain the coal-fired electricity cost within a period corresponding to the coal price forecast result based on the coal price forecast result.

[0174] Furthermore, the central prediction module includes: a model training unit and a central prediction unit;

[0175] The model training unit is used to pre-process the pre-acquired raw data of the second preset area;

[0176] Performing feature screening on the pre-acquired raw data according to a pre-set feature screening algorithm to reduce the feature dimension of the coal price impact feature corresponding to the coal price in the raw data; the coal price impact feature includes at least one feature dimension;

[0177] Based on the original data after feature screening and the pre-set first lag strategy, a corresponding prediction training set is obtained;

[0178] Training a pre-built LSTM network based on the prediction training set to obtain a trained coal price prediction model;

[0179] The central prediction unit is used to input the coal price data group into a preset coal price prediction model when receiving the coal price data group sent by any local data acquisition module, obtain the corresponding coal price prediction result, and send the coal price prediction result to the corresponding local data acquisition module to assist the local data acquisition module in making business decisions.

[0180] Furthermore, when the central prediction module receives a coal price data set sent by any local data acquisition module, it inputs the coal price data set into a pre-set coal price prediction model to obtain a corresponding coal price prediction result, including:

[0181] When a coal price data group is received from any local data acquisition module, a corresponding time coding feature is assigned to the coal price of each unit time in the coal price data group according to the timestamp corresponding to each unit time in the received coal price data group and a preset formula 1; the coal price data group includes the coal price and timestamp corresponding to each unit time in the third preset time period; the formula 1 is:

[0182]

[0183] Wherein, sin(Z) and cos(Z) are the time coding features corresponding to the coal price of any unit time in the coal price data set, Z is the preset time period, and l is the period constant matched with the time period;

[0184] According to the coal price data set and a preset second hysteresis sub-strategy, obtaining a hysteresis feature corresponding to the coal price per unit time in the coal price data set;

[0185] The hysteresis feature is the coal price corresponding to all unit times within the fourth preset period past the unit time corresponding to the coal price at any unit time in the coal price data group;

[0186] According to the preset rolling statistical sub-strategy, the rolling characteristics corresponding to the coal price per unit time in the coal price data group are obtained;

[0187] The rolling characteristics include the coal price of any unit time in the coal price data set, and the mean, standard deviation, maximum value and minimum value of the coal prices corresponding to all unit times in the fifth preset period past the corresponding unit time;

[0188] According to the pre-set differential sub-strategy, the differential features corresponding to the coal price of each unit time in the coal price data set are obtained;

[0189] The differential feature includes the difference between the coal price at any unit time in the coal price data set and the coal price corresponding to the unit time that is six preset times past the unit time.

[0190] Dividing the coal price data group into at least one time window according to a time series according to a preset sliding window;

[0191] According to the pre-set attention mechanism, corresponding attention weights are assigned to the coal prices of all unit times in each time window;

[0192] According to the coal price of each unit time in all time windows, as well as the time coding features, extrapolation features and attention weights corresponding to the coal price, the prediction results of the market coal price corresponding to the coal price data group are obtained.

[0193] This embodiment provides a coal price prediction system based on a multi-feature fusion LSTM network. Due to the integration of a coal price prediction model with multiple feature dimensions, compared with the existing technology, it can effectively assist coal-fired power plants in formulating scientific, reasonable and forward-looking decision-making strategies in a complex and changing market environment, thereby significantly improving the company's dominant position and risk response capabilities in market competition.

[0194] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0195] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; and internal communication between two components or interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0196] In the present invention, unless otherwise expressly specified or limited, when a first feature is "above" or "below" a second feature, it may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Furthermore, when a first feature is "above," "above," or "above" a second feature, it may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is at a higher level than the second feature. When a first feature is "below," "below," or "below" a second feature, it may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is at a lower level than the second feature.

[0197] In the description of this specification, the terms "one embodiment", "some embodiments", "embodiments", "examples", "specific examples" or "some examples" refer to the specific features, structures, materials or characteristics described in conjunction with the embodiment or example and included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and features of different embodiments or examples, unless they are mutually inconsistent.

[0198] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may alter, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A coal price prediction system based on multi-feature fusion LSTM network, characterized by: include: a central prediction module, and at least one local data acquisition module in communication with the central prediction module; Any of the local data acquisition modules is used to obtain a data set of coal prices to be predicted in a first preset area, and send the data set of coal prices to be predicted to the central prediction module; The central prediction module is configured to, upon receiving a coal price data set sent by any local data collection module, input the coal price data set into a pre-set coal price prediction model to obtain a corresponding coal price prediction result, and send the coal price prediction result to the corresponding local data collection module to assist the local data collection module in making business decisions; The coal price prediction model includes: a prediction model obtained by training an LSTM network through a prediction training set; The prediction training set is a training set obtained by performing feature screening on the original data of the second preset area acquired in advance according to a preset feature screening algorithm, and based on the original data after feature screening and a preset first hysteresis strategy; The original data includes the coal price corresponding to each unit time in the first preset time period and the coal price influencing characteristics affecting the change of the coal price; The first hysteresis strategy is to establish a mapping relationship between the coal price corresponding to any unit time in the original data and the historical coal price data corresponding to a second preset time past the unit time.

2. The coal price prediction system based on multi-feature fusion LSTM network according to claim 1 is characterized in that: The central prediction module includes: a model training unit and a central prediction unit; The model training unit is used to pre-process the pre-acquired raw data of the second preset area; Performing feature screening on the pre-acquired raw data according to a pre-set feature screening algorithm to reduce the feature dimension of the coal price impact feature corresponding to the coal price in the raw data; the coal price impact feature includes at least one feature dimension; Based on the original data after feature screening and the pre-set first lag strategy, a corresponding prediction training set is obtained; Training a pre-built LSTM network based on the prediction training set to obtain a trained coal price prediction model; The central prediction unit is used to input the coal price data group into a preset coal price prediction model when receiving the coal price data group sent by any local data acquisition module, obtain the corresponding coal price prediction result, and send the coal price prediction result to the corresponding local data acquisition module to assist the local data acquisition module in making business decisions.

3. The coal price prediction system based on multi-feature fusion LSTM network according to claim 2 is characterized in that: The model training unit pre-processes the pre-acquired raw data of the second preset area, including: performing abnormal data detection on the pre-acquired raw data of the second preset area according to a preset Z-score algorithm, marking abnormal data in the raw data, and replacing the marked abnormal data according to a preset abnormal data replacement algorithm; The abnormal data replacement algorithm includes data difference and / or median replacement.

4. The coal price prediction system based on multi-feature fusion LSTM network according to claim 2 is characterized in that: The model training unit performs feature screening on the pre-acquired raw data according to a pre-set feature screening algorithm, including: According to the preset Pearson correlation analysis algorithm, Lasso regression algorithm, grey relational analysis algorithm and random forest regression algorithm, feature screening is performed on the coal price impact features corresponding to the coal price in the original data; The coal price impact characteristics retained after filtering the original data using Pearson correlation analysis, Lasso regression, grey correlation analysis, and random forest regression were obtained to reduce the characteristic dimensions of the coal price impact characteristics corresponding to the coal prices in the original data.

5. The coal price prediction system based on multi-feature fusion LSTM network according to claim 4 is characterized in that: The model training unit performs feature screening on the coal price impact features corresponding to the coal prices in the original data according to the preset Pearson correlation analysis algorithm, Lasso regression algorithm, grey relational analysis algorithm and random forest regression algorithm, including: Obtaining the correlation coefficient between the coal price influencing feature corresponding to each feature dimension and the coal price based on the original data and a preset Pearson correlation analysis algorithm; Based on the correlation coefficient between the coal price impact feature corresponding to each feature dimension and the coal price and a preset first correlation coefficient threshold, feature screening is performed on the coal price impact feature corresponding to the coal price in the original data: the Pearson correlation analysis includes: Among them, x i is the value of the coal price impact feature of any feature dimension at the i-th unit time, y i is the coal price corresponding to the i-th unit time, is the average value of the coal price impact feature of any characteristic dimension in n unit time, is the average coal price in n unit time, r is the correlation coefficient between the coal price influencing feature of this feature dimension and the coal price; Based on the original data and the preset Lasso regression algorithm, feature screening is performed on the coal price impact features corresponding to the coal price in the original data; the objective function corresponding to the Lasso regression algorithm is: Where y is the corresponding coal price matrix in the original data, X is the feature matrix corresponding to the coal price impact feature of any feature dimension, β is the preset regression coefficient, and α is the preset regularization parameter; Obtaining, based on the original data, a feature sequence corresponding to the coal price impact feature of each feature dimension and a coal price sequence in the original data; Obtaining the correlation between the coal price influencing feature corresponding to each feature dimension and the coal price based on the feature sequence, the coal price sequence and a preset grey correlation analysis algorithm; According to the correlation between the coal price influencing feature corresponding to each feature dimension and the coal price and a preset second correlation threshold, the coal price influencing feature corresponding to the coal price in the original data is subjected to feature screening: the grey correlation analysis algorithm includes: Where c is the correlation between the coal price impact feature corresponding to any feature dimension and the coal price, Δ(x o ,x o ) is the difference between the characteristic series and the coal price series; According to a preset random forest regression algorithm, feature screening is performed on coal price impact features corresponding to the coal prices in the original data.

6. The coal price prediction system based on multi-feature fusion LSTM network according to claim 2 is characterized in that: The model training unit obtains a corresponding prediction training set based on the original data after feature screening and a preset first hysteresis strategy, including: Based on a preset first hysteresis strategy, a mapping relationship is established between the coal price corresponding to any unit time in the original data and the historical coal price data corresponding to a second preset time past the unit time; Obtaining a mapping data set based on a mapping relationship between the coal price corresponding to each unit time in the original data and the historical coal price data corresponding to a second preset time past the unit time; The mapping dataset is divided into a prediction training set and a prediction test set in a ratio of 8:

2.

7. The coal price prediction system based on multi-feature fusion LSTM network according to claim 6 is characterized in that: The model training unit trains the pre-built LSTM network based on the prediction training set to obtain a trained coal price prediction model, including: The pre-built LSTM network is trained based on the prediction training set. During the training process, the corresponding loss function is obtained in real time according to the pre-set Adam optimizer and mean square error algorithm to adjust the hyperparameters of the coal price prediction model in real time to obtain a trained coal price prediction model; Then, the model training unit is further used to: The coal price prediction model is tested based on a pre-set prediction test set to determine whether it is qualified.

8. The coal price prediction system based on multi-feature fusion LSTM network according to claim 1 is characterized in that: The central prediction module, upon receiving a coal price data set sent by any local data acquisition module, inputs the coal price data set into a pre-set coal price prediction model to obtain a corresponding coal price prediction result, including: When a coal price data group is received from any local data acquisition module, a corresponding time coding feature is assigned to the coal price of each unit time in the coal price data group according to the timestamp corresponding to each unit time in the received coal price data group and a preset formula 1; the coal price data group includes the coal price and timestamp corresponding to each unit time in the third preset time period; the formula 1 is: Wherein, sin(Z) and cos(Z) are the time coding features corresponding to the coal price of any unit time in the coal price data set, Z is the preset time period, and l is the period constant matched with the time period; According to the pre-set data extrapolation strategy, the extrapolation characteristics corresponding to the coal price per unit time in the coal price data group are obtained; Dividing the coal price data group into at least one time window according to a time series according to a preset sliding window; According to the pre-set attention mechanism, corresponding attention weights are assigned to the coal prices of all unit times in each time window; According to the coal price of each unit time in all time windows, as well as the time coding features, extrapolation features and attention weights corresponding to the coal price, the prediction results of the market coal price corresponding to the coal price data group are obtained.

9. The coal price prediction system based on multi-feature fusion LSTM network according to claim 8 is characterized in that: The data extrapolation strategy includes a second lag sub-strategy, a rolling statistics sub-strategy and a differential sub-strategy; Then, the central prediction module obtains the extrapolation features corresponding to the coal price per unit time in the coal price data set according to the pre-set data extrapolation strategy, including: According to the coal price data set and a preset second hysteresis sub-strategy, obtaining a hysteresis feature corresponding to the coal price per unit time in the coal price data set; The hysteresis feature is the coal price corresponding to all unit times within the fourth preset period past the unit time corresponding to the coal price at any unit time in the coal price data group; According to the preset rolling statistical sub-strategy, the rolling characteristics corresponding to the coal price per unit time in the coal price data group are obtained; The rolling characteristics include the coal price of any unit time in the coal price data set, and the mean, standard deviation, maximum value and minimum value of the coal prices corresponding to all unit times in the fifth preset period past the corresponding unit time; According to the pre-set differential sub-strategy, the differential features corresponding to the coal price of each unit time in the coal price data set are obtained; The differential feature includes the difference between the coal price at any unit time in the coal price data set and the coal price corresponding to the unit time that is six preset times past the unit time.

10. The coal price prediction system based on multi-feature fusion LSTM network according to claim 1 is characterized in that: Any of the local data acquisition modules is set in a corresponding coal-fired power plant; the first preset area is the coal-fired power plant where the local data acquisition module is set; Any of the local data acquisition modules is further configured to: When the coal price forecast result sent by the central forecast module is received, the coal-fired electricity cost in the time period corresponding to the coal price forecast result is obtained according to the coal price forecast result.

Images