Chemical enterprise profit prediction method based on multi-scale long short-term memory network

Abstract

The invention discloses a chemical enterprise profit prediction method based on a multi-scale long short-term memory network. The chemical enterprise profit prediction method comprises the steps of collecting time series data used for prediction from historical data of a chemical enterprise, preprocessing the collected data, and dividing the processed data into a training set and a test set. And constructing a prediction model based on the multi-scale long and short-term memory network for predicting the financial profit of the chemical enterprise. And respectively training short-term, medium-term and long-term LSTM models by using the data of the training set. And after training is completed, taking the test set as input, obtaining a predicted value of each scale model, calculating a comprehensive error of each model, dynamically adjusting a weighting coefficient of the model according to the comprehensive error, and carrying out weighted fusion on a predicted result according to the weighting coefficient of each model, so as to obtain a final predicted value of the comprehensive profit of the chemical enterprise. The profit time sequence dependency relationship of different time scales can be captured at the same time, prediction is performed by integrating short-term fluctuation and long-term trend information, and the accuracy and robustness of prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis in chemical enterprises, especially for trend prediction of financial data in chemical enterprises, and specifically relates to a profit prediction method for chemical enterprises based on a multi-scale long short-term memory network (Multi-Scale LSTM). Background Art

[0002] During the operation of chemical enterprises, due to the influence of multiple internal and external factors, financial data often exhibits obvious periodic and volatile characteristics. These characteristics include the combined effects of factors such as market demand changes, raw material price fluctuations, production capacity fluctuations, and the macroeconomic cycle, making it very complex to predict the profits of enterprises. In this context, traditional financial prediction methods often struggle to fully reflect the complex non-linear relationships and long-term dependencies in the financial data of chemical enterprises. Therefore, there is an urgent need for a new and more accurate prediction method to address these challenges.

[0003] Most traditional financial prediction methods rely on statistical and econometric models, such as regression analysis, time series analysis (such as the ARIMA model), etc. These methods model the statistical laws of historical data to predict future financial performance. Although in some stable and less-changing environments, these methods can provide relatively reliable predictions, the financial data of chemical enterprises is often affected by multiple factors and has strong periodic and non-linear characteristics. For example, seasonal fluctuations in market demand, periodic changes in raw material prices, and periodic adjustments in policies will all have a profound impact on enterprise profits. Traditional methods are unable to effectively model their non-linear relationships when faced with these complex and intertwined factors, resulting in low prediction accuracy.

[0004] Econometric models, such as the vector autoregressive (VAR) model, attempt to improve prediction accuracy through the relationships between multiple variables. Such models can consider the influence of multiple factors and analyze and predict based on their historical data. However, the financial data of chemical enterprises is usually affected by a large number of variable factors, and the interactions between these factors are not linear but have strong time-lag effects. Traditional econometric methods often struggle to be comprehensive and accurate in dealing with these time-lag effects and non-linear relationships. Therefore, the effects of these methods in practical applications are not ideal.

[0005] In recent years, with the rapid development of deep learning technology, neural network-based models have gradually become important tools for time series prediction. In particular, the Long Short-Term Memory network (LSTM) can effectively solve the problem of gradient disappearance in traditional neural networks due to its unique gating mechanism, making LSTM show superiority in modeling long-term dependencies in time series data. LSTM can provide effective modeling means for short-term and long-term trends by automatically learning long-term dependencies in the data.

[0006] The financial data of chemical enterprises exhibits significant multi-scale characteristics. In the short term, the profit fluctuations of enterprises may be affected by factors such as seasonal demand and raw material price fluctuations; in the long term, factors such as the macroeconomic cycle, industry technological progress, and policies may affect the overall financial performance of enterprises. Introducing a multi-scale LSTM model can effectively capture short-term and long-term fluctuations in the financial data of chemical enterprises by processing information at multiple time scales simultaneously. However, traditional model fusion methods have limitations in dealing with the financial data of chemical enterprises, mainly reflected in the fact that the weighted coefficients of the models are difficult to adapt to the dynamic changes in the actual situation, so they cannot make good use of the information captured by the multi-scale LSTM model at different time scales, resulting in poor accuracy and stability of model prediction. Summary of the Invention

[0007] 1. Technical problems to be solved by the invention

[0008] In view of the fact that the financial data of chemical enterprises is usually affected by various factors, showing obvious periodic fluctuations and non-linear characteristics, traditional financial prediction methods are difficult to accurately capture these complex laws. The present invention provides a profit prediction method for chemical enterprises based on a multi-scale long short-term memory network. The present invention aims to improve the accuracy and robustness of prediction through comprehensive analysis of multi-time scale data, so as to better support the financial decision-making and strategic planning of enterprises.

[0009] 2. Technical solution

[0010] To achieve the above object, the technical solution provided by the present invention is as follows:

[0011] A profit prediction method for chemical enterprises based on a multi-scale long short-term memory network of the present invention includes the following steps:

[0012] S1. Data collection and preprocessing: Collect time series data for prediction from the historical data of chemical enterprises, and sequentially perform cleaning and standardization operations on the collected data, and divide the processed data into a training set and a test set;

[0013] S2. Model construction: Construct a financial profit prediction model for chemical enterprises based on a multi-scale long short-term memory network. The model includes a short-term LSTM model, a medium-term LSTM model, and a long-term LSTM model. Models with different time scales process data with different time scales respectively;

[0014] S3. Model training: Use the data in the training set to train the short-term LSTM model, the medium-term LSTM model, and the long-term LSTM model respectively, and capture small-scale, medium-scale, and large-scale feature information in the profit data;

[0015] S4. Model fusion: Input the test set into models with different time scales to obtain predicted values, calculate the comprehensive error and cumulative comprehensive error of each model, dynamically adjust the weighted coefficients of the models according to the cumulative comprehensive error, and perform weighted fusion on the prediction results according to the weighted coefficients to obtain the final predicted value;

[0016] S5. Model evaluation and optimization: Calculate error metrics for the prediction results after weighted fusion, and optimize the hyperparameters of the LSTM model based on error analysis to improve the prediction accuracy.

[0017] Furthermore, the historical data collected in step S1 includes: sales volume, production cost, raw material price, product market price, inventory level, and historical daily profit; After preprocessing the collected historical data, it is input into the prediction model constructed in step S2, and the output target of the model is the predicted value of future daily profit.

[0018] Furthermore, after cleaning and standardizing the collected data in step S1, the sliding window method is used to divide the data into sample sets with three time scales: short-term s days, medium-term m days, and long-term l days. Each sample set contains the data of the first T time points as features, and the data of the next time point as labels, and the training set and test set are divided according to a certain proportion.

[0019] Furthermore, the short-term LSTM model, the medium-term LSTM model, and the long-term LSTM model constructed in step S2 have the same structure, all including an input layer, several hidden layers, and an output layer. Among them, the number of hidden layers of the short-term LSTM model is set to 3, the number of hidden layers of the medium-term LSTM model is set to 2, and the number of hidden layers of the long-term LSTM model is set to 1.

[0020] Furthermore, in step S3, the short-term LSTM model uses a sliding window of s days to process data for training, the medium-term LSTM model uses a sliding window of m days to process data for training, and the long-term LSTM model uses a sliding window of l days to process data for training.

[0021] Further, the weighted coefficient dynamic adjustment mechanism in step S4 is as follows: calculate the comprehensive error E of each model by the predicted value and the true value of each model comi (t), and then based on the comprehensive error E comi (t), calculate the cumulative comprehensive error E of each model through the cumulative comprehensive error formula sumi (t), and calculate the weighted coefficient at the next moment according to the weighted coefficient formula, set the threshold θ to control the weighted coefficient update frequency, and when the cumulative error change ΔE total (t) ≥ θ, update the weighted coefficient.

[0022] Further, the threshold θ is adaptively adjusted according to the degree of data fluctuation and the stability requirements of the model. When the data fluctuates greatly, the threshold is increased, and when the data is stable, the threshold is decreased.

[0023] 3. Beneficial effects

[0024] Adopting the technical solution provided by the present invention, compared with the existing well-known technologies, it has the following remarkable effects:

[0025] (1) A profit prediction method for chemical enterprises based on a multi-scale long short-term memory network of the present invention can capture the profit time series dependence relationships at different time scales simultaneously by constructing a multi-scale LSTM model and setting hidden layers with different jump scales, and make predictions by integrating short-term fluctuations and long-term trend information. Compared with traditional single-scale models and traditional statistical models, it can more accurately reflect data changes.

[0026] (2) A profit prediction method for chemical enterprises based on a multi-scale long short-term memory network of the present invention determines the weighted coefficient by calculating the comprehensive error of each scale model, and performs weighted fusion on the prediction results to make the final predicted value more reasonable. Considering the characteristic that the weighted coefficient of each model should be adjusted with the dynamic change of the cumulative comprehensive error of the model, a mechanism for dynamically adjusting the smoothing coefficient based on the change rate of the cumulative comprehensive error is proposed. As the smoothing coefficient changes, the cumulative comprehensive error of the model will also change accordingly, and then the weighted coefficient of each model is adjusted. And this method sets the weighted coefficient update condition and adjusts the threshold according to the degree of fluctuation of the financial data of chemical enterprises. When the data fluctuation is relatively stable, the threshold is appropriately increased to reduce unnecessary weighted coefficient updates and maintain the stability of the model.

[0027] (3) A profit prediction method for chemical enterprises based on a multi-scale long short-term memory network of the present invention is based on the LSTM model and its multi-scale variants of deep learning, and uses its multi-level non-linear modeling ability to better capture the complex dependence relationships in the data, which is difficult for traditional linear regression or time series analysis methods to achieve, thereby improving the accuracy and robustness of the prediction. Description of the drawings

[0028] Figure 1 Schematic flowchart of a profit prediction method for chemical enterprises based on a multi-scale long short-term memory network according to the present invention;

[0029] Figure 2 Schematic structural diagram of the LSTM model involved in the present invention;

[0030] Figure 3 Fusion prediction effect diagram of a profit prediction method for chemical enterprises based on a multi-scale long short-term memory network according to the present invention. Detailed implementation manners

[0031] To further understand the content of the present invention, the present invention will be described in detail in conjunction with the accompanying drawings and embodiments.

[0032] Embodiment 1

[0033] The financial data of chemical enterprises is usually affected by various factors, showing obvious periodic fluctuations and non-linear characteristics. Traditional financial prediction methods are difficult to accurately capture these complex laws. Existing statistical models such as regression analysis and ARIMA models have low accuracy in dealing with short-term fluctuations and long-term dependence relationships, and cannot fully cope with the influence of factors such as changing market demands and raw material price fluctuations.

[0034] In addition, although the traditional LSTM model can capture long-term dependence relationships, due to its single time scale limitation, it cannot effectively handle multi-scale fluctuations in the financial data of chemical enterprises. Therefore, how to combine multi-scale data to accurately predict the profit changes of enterprises has become an urgent problem to be solved. This embodiment proposes a profit prediction method for chemical enterprises based on a multi-scale LSTM model, aiming to improve the accuracy and robustness of prediction through comprehensive analysis of multi-time scale data, so as to better support the financial decision-making and strategic planning of enterprises.

[0035] The multi-scale LSTM model proposed in this embodiment constructs multiple LSTM networks in parallel, processes data of different time scales respectively, and combines this information for comprehensive prediction. The internal structure of the LSTM model is as Figure 2 shown. This method can effectively fuse short-term fluctuations and long-term trends, and improve the prediction accuracy of complex financial data. By introducing data inputs of different time scales into the model, the multi-scale LSTM can capture different frequency change patterns in the data, and then improve the modeling ability for periodic fluctuations and long-term dependence relationships. Especially when facing data with obvious periodic fluctuations such as chemical enterprises, the advantages of the multi-scale LSTM are more obvious.

[0036] In addition, the financial data of chemical enterprises is affected not only by time factors but also by other external factors such as raw material prices and changes in market demand. The interaction of these factors is usually non-linear, and traditional linear regression or time series analysis methods are difficult to effectively model these complex non-linear relationships. Based on LSTM and its multi-scale variants, this embodiment better captures the complex dependencies in the data through multi-level non-linear modeling.

[0037] Combined with Figure 1 , the specific steps of this embodiment are as follows:

[0038] S1. Data collection and preprocessing: Collect the original time series data for profit prediction from the historical data of chemical enterprises, and sequentially perform data cleaning and standardization on the collected data. Data cleaning is used to remove missing values and outliers, and the Z-score standardization method is used to standardize the data to ensure that the data can be compared under the same dimension. The Z-score standardization formula is as follows:

[0039]

[0040] where x is the data point, μ is the data mean, and σ is the data standard deviation. Through standardization, the data scale difference can be effectively eliminated.

[0041] In this embodiment, the original data collected from the historical data of chemical enterprises for profit prediction includes: sales volume, production cost, raw material price, product market price, inventory level, and historical daily profit. After preprocessing the collected original data, it is input into the prediction model, and the target data output by the model is the predicted value of future daily profit.

[0042] After preprocessing the data, this embodiment uses the sliding window method to divide the data into sample sets of three time scales: short term (s days), medium term (m days), and long term (l days). Each sample set contains the data of the first T time points as features, and the data of the next time point as the label. The sample set is divided into a training set and a test set in a ratio of 8:2, that is, the first 80% is the training set, and the last 20% is the test set.

[0043] S2. Model construction: Construct a prediction model for the financial profit of chemical enterprises based on a multi-scale long short-term memory network to predict the financial profit of chemical enterprises.

[0044] The multi-scale LSTM model constructed in this embodiment includes a short-term LSTM model, a medium-term LSTM model, and a long-term LSTM model, and each model processes data at different time scales. The short-term LSTM model uses a sliding window of s days to process data for training, the medium-term LSTM model uses a sliding window of m days to process data for training, and the long-term LSTM model uses a sliding window of l days to process data for training.

[0045] Each LSTM model has the same structure, including an input layer, several hidden layers (each layer has 128 neurons), and an output layer. Among them, the number of hidden layers of the short-term LSTM model is set to 3, the number of hidden layers of the medium-term LSTM model is set to 2, and the number of hidden layers of the long-term LSTM model is set to 1. The three models respectively capture the time-series dependence relationship of the profits of chemical enterprises in units of s days, m days, and l days.

[0046] S3. Model training: Sequentially input the preprocessed data as the input data for predicting the profits of chemical enterprises into the short-term LSTM model, the medium-term LSTM model, and the long-term LSTM model, train the short-term LSTM model, the medium-term LSTM model, and the long-term LSTM model, and sequentially capture the small-scale, medium-scale, and large-scale feature information in the profit data.

[0047] S4. Model fusion: After the model training is completed, use the test set as the input and input it into the short-term LSTM model, the medium-term LSTM model, and the long-term LSTM model respectively to obtain the predicted values of each scale model. Calculate the comprehensive error of each model through the predicted values and the true values of each model. The calculation formula for the comprehensive error is:

[0048]

[0049] Among them, y t,k is the true value, is the predicted value at time t, is the comprehensive error, n is the number of currently input samples, k is the sample number, and i is the model identifier of different scales.

[0050] Calculate the cumulative comprehensive error of each model through the comprehensive error of each model. The calculation formula for the cumulative comprehensive error is:

[0051]

[0052] Among them is the comprehensive error at the current moment calculated based on the new predicted value, is the cumulative comprehensive error at the current moment, is the cumulative comprehensive error at the previous moment, and α tis the smoothing coefficient. It can be seen from the formula that the comprehensive cumulative error will be updated over time.

[0053] The calculation formula for the smoothing coefficient is:

[0054]

[0055] where ΔE i (t) is the error change rate, and α t is the smoothing coefficient. In this way, when the error change rate is large, the smoothing coefficient will be closer to 1, making the influence of the historical cumulative synthesis greater; when the error change rate is small, the smoothing coefficient will be closer to 0.5, giving similar smoothing coefficients to the historical cumulative error and the current comprehensive error.

[0056] According to the cumulative comprehensive error of each model, calculate the weighted coefficient of the model. The calculation formula for the weighted coefficient is:

[0057]

[0058] where w i (t + 1) is the weighted coefficient of each model at the next moment, is the cumulative comprehensive error at the current moment.

[0059] To ensure the stability of the model, a condition for updating the weighted coefficient is set. Whether to update the weighted coefficient is determined according to the change amount of the error. The calculation formula for the cumulative error change amount is:

[0060]

[0061] where ΔE total (t) is the cumulative error change amount, is the cumulative error at the current moment, is the cumulative error at the previous moment. When ΔE total (t) ≥ θ, the weighted coefficient is updated, where θ is the set threshold and is adjusted according to the degree of data fluctuation and the stability requirements of the model. If the data itself fluctuates greatly, then the error change amount may also be relatively large. In this case, the threshold θ can be set larger to avoid frequent updating of the weighted coefficient due to normal data fluctuations. On the contrary, if the data is relatively stable, the threshold θ can be set smaller to capture the data changes in time and update the weighted coefficient.

[0062] According to the weighted coefficients of each LSTM model, weight and fuse their prediction results to obtain the final predicted value of the comprehensive profit of the chemical enterprise. The fusion formula is:

[0063]

[0064] where, is the fused prediction value, i.e., the final prediction value, is the prediction value of each scale model at time t+1.

[0065] It should be emphasized that this embodiment proposes a fusion method based on error feedback for multi-scale long short-term memory networks. This method dynamically adjusts the weighting coefficients according to the comprehensive cumulative error situation of each model based on error feedback. When the prediction error of a certain scale model at a certain moment is small, it indicates that this scale model captures the financial data characteristics more accurately under the current situation. At this time, its weighting coefficient can be increased accordingly to make it play a greater role in the fusion process; conversely, when the prediction error of a certain scale model is large, its weighting coefficient is reduced. Through this adaptive weighting coefficient adjustment mechanism, the information captured by the multi-scale LSTM model at different time scales can be better comprehensively utilized, improving the accuracy and stability of the model prediction, and thus more effectively coping with the multi-scale characteristics and complex and changeable influencing factors of chemical enterprise financial data.

[0066] S5. Model evaluation and optimization: Evaluate the prediction results after weighted fusion, calculate the prediction error using indicators such as mean absolute error (MAE) and mean absolute percentage error (MAPE), and optimize the hyperparameters of the LSTM model in a Bayesian optimization manner based on error analysis to improve the prediction accuracy.

[0067] Embodiment 2

[0068] The data is collected from the historical financial data of a chemical enterprise, and the time range is from January 1, 2018 to December 31, 2023. A total of 2190 daily data are collected, including 6 fields such as sales volume, production cost, raw material price, product market price, inventory level, and daily profit. Part of the data is shown in Table 1:

[0069] Table 1 Part of the historical financial data of a chemical enterprise

[0070]

[0071] The data cleaning steps include removing missing values and outliers. The standardization process uses the Z-score standardization method to standardize all fields to ensure that the data can be compared under the same dimension. The sliding window method is used to divide the data into sample sets of three time scales: short-term (1 day), medium-term (7 days), and long-term (30 days). Each sample set contains the data of the first T time points as features and the data of the next time point as labels. The sample sets are divided into training sets and test sets in a ratio of 8:2, that is, the first 80% (1752 data) is the training set, and the last 20% (438 data) is the test set.

[0072] Construct a multi-scale LSTM model, where each model processes data at different time scales. The short-term LSTM model processes data using a 1-day sliding window, the medium-term LSTM model processes data using a 7-day sliding window, and the long-term LSTM model processes data using a 30-day sliding window, and each model is trained.

[0073] After the model training is completed, input the test set data into the short-term, medium-term, and long-term LSTM models to obtain the prediction results of each model respectively. By comparing the predicted values and the actual values of each model, calculate the error of each model, and use these errors to calculate the weighted coefficients of the models. Perform a weighted average on the prediction results of each model according to the weighted coefficients to obtain the final comprehensive profit prediction value. The effect diagram of the test set prediction fusion is as Figure 3 shown. From Figure 3 it can be seen that the profit prediction trend of this method and the actual profit trend have a high similarity, indicating that this method has a good prediction effect and can more accurately reflect the changes in the profits of chemical enterprises.

[0074] Evaluate the prediction results after weighted fusion, and calculate the prediction error using indicators such as the mean absolute error and the mean absolute percentage error. Based on the error analysis, use the Bayesian optimization method to optimize the hyperparameters of the LSTM model. The main hyperparameters to be optimized include the number of neurons in the hidden layer, the learning rate, the regularization parameter, etc. The optimized model can significantly improve the prediction accuracy.

[0075] The above schematically describes the present invention and its implementation manners. This description is not restrictive, and what is shown in the drawings is only one of the implementation manners of the present invention. The actual structure is not limited thereto. Therefore, if those of ordinary skill in the art are inspired by it and design a structural manner and an embodiment similar to the technical solution without creative work without departing from the purpose of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A profit prediction method for chemical enterprises based on a multi-scale long short-term memory network, characterized in that It includes the following steps: S1. Data collection and preprocessing: Collect time series data for prediction from the historical data of chemical enterprises, and successively perform cleaning and standardization operations on the collected data, and divide the processed data into a training set and a test set; S2. Model construction: Construct a financial profit prediction model for chemical enterprises based on a multi-scale long short-term memory network. The model includes a short-term LSTM model, a medium-term LSTM model, and a long-term LSTM model. Models with different time scales process data with different time scales respectively; S3. Model training: Use the data in the training set to train the short-term LSTM model, the medium-term LSTM model, and the long-term LSTM model respectively to capture small-scale, medium-scale, and large-scale feature information in the profit data; S4. Model fusion: Input the test set into models with different time scales to obtain predicted values, calculate the comprehensive error and cumulative comprehensive error of each model, dynamically adjust the weighted coefficients of the models according to the cumulative comprehensive error, and perform weighted fusion on the prediction results according to the weighted coefficients to obtain the final predicted value; S5. Model evaluation and optimization: Calculate error indicators for the prediction results after weighted fusion, and optimize the hyperparameters of the LSTM model based on error analysis to improve the prediction accuracy.

2. The profit prediction method for chemical enterprises based on multi-scale long short-term memory network according to claim 1, characterized in that: The historical data collected in step S1 includes: sales volume, production cost, raw material price, product market price, inventory level, and historical daily profit; after preprocessing the collected historical data, input it into the prediction model constructed in step S2, and the model output target is the predicted value of future daily profit.

3. A profit prediction method for chemical enterprises based on a multi-scale long short-term memory network according to claim 1 or 2, characterized in that: After performing cleaning and standardization operations on the collected data in step S1, the sliding window method is used to divide the data into sample sets with three time scales: short-term s days, medium-term m days, and long-term l days. Each sample set contains the data of the first T time points as features, and the data of the next time point as labels, and divide the training set and the test set according to a ratio.

4. A profit prediction method for chemical enterprises based on a multi-scale long short-term memory network according to claim 3, characterized in that: The short-term LSTM model, medium-term LSTM model, and long-term LSTM model constructed in step S2 have the same structure, all including an input layer, several hidden layers, and an output layer. Among them, the number of hidden layers of the short-term LSTM model is set to 3, the number of hidden layers of the medium-term LSTM model is set to 2, and the number of hidden layers of the long-term LSTM model is set to 1.

5. The profit prediction method for chemical enterprises based on multi-scale long short-term memory network according to claim 4, wherein: In step S3, the short-term LSTM model uses a sliding window of s days to process data for training, the medium-term LSTM model uses a sliding window of m days to process data for training, and the long-term LSTM model uses a sliding window of l days to process data for training.

6. The profit prediction method for chemical enterprises based on a multi-scale long short-term memory network according to claim 4, characterized in that: The weighted coefficient dynamic adjustment mechanism in step S4 is as follows: calculate the comprehensive error of each model by using the predicted value and the true value of each model respectively Then, based on the comprehensive error Calculate the cumulative comprehensive error of each model through the cumulative comprehensive error formula And calculate the weighted coefficient at the next moment according to the weighted coefficient formula, and set a threshold θ to control the update frequency of the weighted coefficient. When the change amount of the cumulative error ΔE total (t) ≥ θ, update the weighted coefficient.

7. The profit prediction method for chemical enterprises based on multi-scale long short-term memory network according to claim 6, characterized in that: The threshold θ is adaptively adjusted according to the fluctuation degree of the data and the stability requirements of the model. When the data fluctuates greatly, the threshold is increased, and when the data is stable, the threshold is decreased.

8. A profit prediction method for chemical enterprises based on a multi-scale long short-term memory network according to claim 6, characterized in that: Comprehensive error The calculation formula is as follows: where y t,k is the true value, is the predicted value at time t, n is the number of currently input samples, k is the sample number, and i is the model identifier for different scales; The calculation formula for the cumulative comprehensive error is: Among them, is the current comprehensive error calculated based on the new predicted value, is the cumulative comprehensive error at the current moment, is the cumulative comprehensive error at the previous moment, and α t is the smoothing coefficient; The calculation formula for the weighted coefficient is: Among them, w i (t + 1) is the weighted coefficient of each model at the next moment, is the cumulative comprehensive error at the current moment.

9. The profit prediction method for chemical enterprises based on multi-scale long short-term memory network according to claim 8, characterized in that: The calculation formula for the change amount of the cumulative error is: where ΔE total (t) is the cumulative error change amount, is the cumulative error at the current moment, is the cumulative error at the previous moment.

10. A profit prediction method for chemical enterprises based on a multi-scale long short-term memory network according to claim 8, characterized in that: The formula for weighted fusion of the prediction results according to the weighted coefficient in step S4 is: Among them, is the fusion prediction value, that is, the final prediction value, is the prediction value of each scale model at time t+1.

Images