Enterprise power emission reduction technology based on the LSTM–SVM model

3.1 The law and classification of electricity load in enterprises

The concept of energy conservation and emission reduction is becoming increasingly sophisticated in major enterprises. Electricity is an indispensable component of many industries and has become an important goal for enterprises to reduce emissions. The objective of analyzing electricity data for emission reduction is to achieve three key outcomes: the timely prediction of electricity consumption data, the provision of early warning for peak electricity, and the establishment of a reference significance for subsequent electricity consumption plans of enterprises in the form of data. These results collectively contribute to the control of electricity consumption. The electricity consumption patterns and loads vary among different enterprises, such as large industrial enterprises, which usually require their machines to operate with electricity for a long time. This also leads to poor regularity and planning of electricity consumption in industrial enterprises, making it more difficult for enterprises to provide electricity warnings. Therefore, this study takes a certain steel enterprise as an example to analyze the composition and patterns of its electricity consumption load and divides it into two categories: shock load and stable load. After predicting them separately, they are fused to achieve data warning. The visualization curves of different loads are shown in Figure 1.

Figure 1

Comparison of shock load and stable load curves: (a) impact load and (b) steady load.

The data in Figure 1 are from the monitoring and statistics of power load in a certain area. Electricity needs to be continuously supplied in steel enterprises, which requires that electricity can be predicted in the short term. Short-term predictions are easily influenced by multiple factors, leading to sudden declines or increases. Consequently, when formulating projections regarding future data, it is imperative to consider not only the prospective trajectory but also to integrate historical data and assess the influence of production under disparate seasonal, climatic, and other conditions. This study first analyzes the shock load, but due to its randomness and time-varying nature, it cannot be directly described through mathematical models. Therefore, this study chooses the covariance correlation coefficient method for calculation and further analyzes the factors related to electricity consumption characteristics, as shown in the following equation [16,17]:

In Eq. (1), L i = { l i ( 1 ) , l i ( 2 ) , … , l i ( k ) , … , l i ( n ) } represents the set of loads. P i = { p i ( 1 ) , p i ( 2 ) , … , p i ( k ) , … , p i ( n ) } represents the set of steel prices that affect the load. i represents the number of days. k / n represent the sampling interval and total number of days, respectively, E is the mathematical expectation. The variance of different sets is shown in the following equation:

In Eq. (2), D ( L i ) / D ( P i ) represent the variance of the load set and the steel price set. The covariance coefficients ρ L i , P i of the two sets are shown in the following equation:

Excluding the steel price factor, electricity prices, climate, and other factors can all have an impact on the electricity load, and the same method should be used for calculation. To ensure the accuracy of the calculation results, this study introduces the Pearson algorithm and conducts correlation analysis to ensure calculation accuracy. Abundant data indicate that the impact of electricity load on enterprises is more closely related to steel prices, temperature, humidity, and day types. The prediction difficulty of stable load is relatively low compared to shock load. The results of this study using the same method for operational analysis indicate that the correlation between stable electricity load and steel price, daily type, electricity price, and temperature is higher. Based on the above analysis, the shock load input and stable load input of the model can be obtained, as shown in Figure 2.

Figure 2

Shock load input and stable load input of the model.

In Figure 2, the historical data in the shock load include the simultaneous load for the first 30 days of the prediction and the load at each time point for the first 24 h of the prediction. The day types of shock load are classified according to working days and holidays, represented by 1 and 0, respectively. The historical load data of stable load refer to the data from the week before the forecast. The day type also represents the data from the week before the forecast, with the same symbol as the shock load (Table 1).

3.2 Construction of a basic PDP system based on deep learning

This study chooses LSTM for model construction. An LSTM network represents a variant of an RNN. It combines the memory storage function of traditional RNN models with a new long-term and short-term memory function. The LSTM model optimizes the long-term dependency phenomenon of the RNN model through input gates, forget gates, etc. The “gate” aims to control the state of memory, storage time, and distance data, and its overall structure is shown in Figure 3.

Figure 3

LSTM structure.

In Figure 3, the input gate i t , output gate o t , and forget gate f t of the model are shown in the following equation [18]:

In Eq. (4), W i / W o / W f and b i / b o / b f represent the weight matrix and offset of the corresponding “gate” at time t . x t / y t / s t represent the input and output values of the model at time t , as well as the memory state of the neuron. The candidate state of neurons is related to x t / y t − 1 , as shown in the following equation:

In Eq. (5), W s / b s represent the weight matrix and offset of the candidate states. The overall operation process of LSTM is shown in Figure 4.

Figure 4

LSTM operation process analysis.

Figure 4 illustrates the preprocessing stage, which aims to attribute impact and stabilize the load data from 0 to 1. Subsequently, the neurons are activated, a process associated with the weight parameter, specifically weight ( w t , b t ) and offset ( w t , b t ) at time t . The system state of the neurons is shown in the following equation:

In Eq. (6), v i represents the input to the neuron. θ represents the weight value size of the cycling cell. h j represents the system state of the neuron. a i indicates the bias. The weight parameters at time t are associated with the information from the previous time. Subsequently, the system is trained based on the output value from the previous moment and the input value from the current moment, with the objective of outputting parameters. The squared reconstruction error is employed as the cost function. Once the training phase is complete, the predicted output data undergo further normalization to obtain the true load value of the system. The calculation of the discrepancy between the predicted and actual values can be utilized to assess the precision of the model’s forecasting capabilities.

3.3 Design of a combined EDEW system based on SVM optimization

SVM is a generalized linear classifier that performs binary classification on data through supervised learning. It conducts machine learning by minimizing structural risk, and the learning strategy is to maximize the interval. The SVM model meets the requirements of small sample learning and has high robustness, which compensates for the large sample demand of LSTM systems. Combined with structural risk minimization and fitting error, the objective function R ( w ) of the model can be obtained, as shown in the following equation [19]:

In Eq. (7)， w represent the weight coefficient normal vector. φ ( x ) indicates the mapping function. b represents an arbitrary constant. C indicates the penalty parameters. ε represents the insensitive loss function. Then, according to the Lagrangian, the regression function of the model can be obtained as shown in the following Eq. (8):

In Eq. (8), α i − α i ⁎ represents the Lagrange coefficient and K ( x i , x ) represents the kernel function. The load prediction can be treated as a multi-classification problem, with the appropriate multi-classifier constructed either by the direct or by the indirect method. The influencing factors present at a given time can be regarded as the basis for classification, with the predicted value of the load at that time representing the label for the final classification result. The primary prediction process is as follows: initially, the data undergo processing and normalization, with the input data adopting the same preprocessing method. Subsequently, the optimal penalty parameter, kernel function, and insensitive loss function are identified through the application of the test-finding method. Then, the input parameters are used for simulation testing, and the data are reverse-normalized to obtain the predicted values of the data. Finally, the resulting load prediction value is compared with the actual load, and the relative error is used as a measure to measure the prediction accuracy of the SVM model. The dimension of the feature space needs to meet the requirements, as shown in Figure 5.

Figure 5

Transformation of nonlinear problems in the SVM model.

Integrating the LSTM model and the SVM model can achieve complementary independent models. For example, the LSTM model compensates for the temporal defects of the SVM model, while the SVM model compensates for the shortcomings of the LSTM model in sample requirements. This study introduces a fuzzy logic algorithm for the weight allocation of each model. The principle is to calculate the difference between the predicted value of the model and the estimated value from the previous moment and select the allocation situation with the lowest estimation deviation. The predicted value of the shock load is shown in the following equation [20,21,22,23]:

In Eq. (9), Y iron _ LSTM / Y iron _ SVM represent the estimated shock load values of LSTM and SVM models. ω 1 j / ω 2 j represent the corresponding model weights. The weight calculation of the LSTM is shown in the following equation:

In Eq. (10), K 1 j = ∣ Y j iron _ LSTM − Y j − 1 ∣ and K 2 j = ∣ Y j iron _ SVM − Y j − 1 ∣ represent the kernel functions of LSTM and SVM models, respectively. Y j − 1 represents the estimated value of the load data from the previous moment. x j is the variable associated with the time sampling point. The weight of the SVM model is the difference between the weight of 1 and the weight of the LSTM model. The predicted value of stable load is shown in the following equation:

In Eq. (11), Y stable _ LSTM / Y stable _ SVM represent the estimated stable load values of LSTM and SVM, respectively. In summary, the combined EDEW system based on LSTM–SVM is shown in Figure 6.

Figure 6

LSTM–SVM combined EDEW system.

First, the data need to be preprocessed separately in the LSTM model and SVM model, that is normalization. The number of hidden layers and neurons in LSTM directly affects the training speed and final accuracy. An increase in the number of hidden layers and neurons results in greater complexity of the model, which may not necessarily lead to enhanced accuracy of the network. It is found that the prediction accuracy improves significantly with the increase of layers, but it is also prone to overfitting when the number of layers is too high. The SVM model is adjusted by the test method. The parameters include the penalized parameter best value, the kernel function, and the insensitive loss function. To avoid overfitting and underfitting the model, the study uses the K-CV method to train the sample data. The trial method is employed to enable the parameters and penalty factors to ascertain the optimal range, after which a small-range test is conducted. After multiple trials, the value is determined [2,3,4]. Finally, the prediction results of the two models are calculated through the weight allocation, and the final predicted value is output. The weight allocation in the LSTM–SVM integrated model is based on the ratio of the difference between the LSTM and SVM algorithms and the predicted value of the previous moment. This weight can be used to identify the inflection point of the present moment and the previous moment. The study obtains the load prediction results of the combined algorithm by calculating the weights of all samples at the prediction time. When the inflection point is large, the trusted SVM algorithm predicts the results, while the LSTM algorithm is the main one.

In the data preprocessing module of the system, the input is raw impact load and stable load data, which are normalized and output into a unified format of data. The input of the parameter tuning module is preprocessed data, and the parameters of the LSTM and SVM models are determined through experimental methods to output the optimal parameter combination. The training module uses optimized parameters and data to train LSTM and SVM models, respectively, and outputs the trained models. The simulation testing module takes the trained model and test data as inputs and outputs the prediction results of the two models. Finally, the prediction results are calculated through weight allocation, and the final prediction value of the combined model is output. The weights are obtained based on the weights of all samples within the prediction time.

4.1 Performance optimization effect based on the LSTM–SVM model

To verify the effectiveness of the research algorithm, experimental analysis is first conducted on the performance of the model itself. Table 2 shows the specific experimental environment parameters.

This study first takes the electricity data of a steel manufacturing plant for the past week as an example to analyze the model load prediction error. The performance comparison of the model before and after combination is shown in Figure 7.

Figure 7

Comparison of model prediction errors before and after optimization: (a) relative error variation line and (b) relative error of each model.

Figure 7(a) shows the variation curve of the relative prediction error (RPE) of the model before and after optimization. The RPE of both LSTM and SVM models before the combination is greater than 1%. Among them, the independent SVM model has the worst prediction performance, with its lowest RPE reaching 2.06%. Compared to the LSTM–SVM model, the minimum RPE is 0.20%, which is 1.86% higher. The RPE fluctuation of LSTM–SVM is also the smallest, with a maximum fluctuation of 0.78%. Compared to LSTM and SVM, the maximum RPE fluctuations are reduced by 1.55 and 1.04%, respectively. In Figure 7(b), this study specifically compares the maximum RPE and average RPE of each model. LSTM–SVM has significant differences compared to other models in both indicators. The maximum RPE of the research model is also less than 1%, a decrease of 2.64% compared to the other two models. The average RPE of the research model is 0.61%, a decrease of 2.11% compared to the other two models. In summary, the LSTM–SVM model that has been optimized through combination has significantly improved prediction performance compared to before optimization.

The study then selects the bidirectional LSTM network (BiLSTM), gated recurrent unit (GRU), bidirectional GRU (BiGRU), and DeepAR probabilistic forecasting algorithm (DeepAR) as alternative algorithms to the LSTM–SVM model. These algorithms are trained and tested on the same historical power dataset, which includes information such as power load during different production periods and under various equipment operating conditions. This is done to evaluate the performance of the LSTM–SVM model, with the results presented in Table 3.

As shown in Table 3, the LSTM–SVM model demonstrates significant advantages across all evaluation metrics. It has the lowest mean square error (MSE) and mean absolute error (MAE), indicating that the prediction values are closer to the actual values, thereby offering higher accuracy. The coefficient of determination (R ²) is closest to 1, suggesting that the model fits the data well and effectively explains the patterns in power data. In comparison, while BiLSTM, GRU, BiGRU, and DeepAR also perform well, they fall slightly short in terms of precision and fit. This indicates that the LSTM–SVM model combines the advantages of LSTM in processing sequence data with the classification and regression capabilities of SVM and can more effectively capture complex features and patterns in power data. Consequently, it offers better performance and applicability in power data early warning for corporate emission reduction.

This study further compares the precision–recall (PR) and ROC curves of the LSTM model and the combination model, and the results are shown in Figure 8.

Figure 8

Performance comparison between the LSTM model and the combined model: (a) ours ROC curve, (b) LSTM model ROC curve, (c) ours PR curve, and (d) LSTM model PR curve.

Figure 8(a) and (b) show the ROC curves of the study model and LSTM, respectively, showing the balance between sensitivity and false negative probability for each model. Therefore, the area under the curve (AUC) composed of the curve and the horizontal axis can evaluate the authenticity of the model. The AUC of the research model reaches 0.94, which is 12.77% higher than the LSTM model’s 0.82. At the equilibrium point, the sensitivity and false negative classes of the research model are 96.98 and 21.42%, respectively, with a relative increase of 11.49% in sensitivity and a relative decrease of 9.87% in false negative. Figure 8(c) and (d) show the comparison of PR curves between the study model and the LSTM model. At the equilibrium point, the accuracy of the research model reaches 95.73%. Compared to the LSTM model, it has improved by 12.85%. The recall rate of the research model reaches 97.96%, which is 11.60% higher than LSTM. Therefore, the research model has shown significant improvements in accuracy, recall, and sensitivity.

4.2 Performance comparison and practical application analysis of the LSTM–SVM model

To further validate the performance of the model in the overall study, this study compares the data prediction model based on ASC and LSTM networks (ASC-LSTM) proposed by Guo et al. with the MTL optimization model proposed by Liu et al. Table 4 shows the results of multiple tests conducted at different voltage levels.

In Table 4, when the voltage level is 35 kV, the differences in indicators among the models are relatively small. However, as the voltage level increases, the model differences will also gradually increase. In terms of training time indicators, the three models have very similar training times under 35 kV conditions, with a difference of less than 3%. For voltage levels of 110 and 220 kV, the training time of the research model is still the shortest, demonstrating its efficiency in processing higher dimensional data. The average training time of the research model is 35.3 ms, while the average of ASC-LSTM is 37.3 ms, and the MTL is 36.5 ms. Therefore, in terms of training time, the research model has relatively reduced by 4.33%. The testing time of each model still maintains the same pattern. The average testing time for the research model is 15.0 ms, while ASC-LSTM is 17.4 ms and MTL is 20.3 ms. Therefore, in the testing time indicator, the research model has relatively decreased by 19.95%. In the prediction accuracy indicators, at the 35 kV voltage level, the accuracy of the three models is relatively high, and the difference is small. At the 110 kV voltage level, the accuracy of the research model is significantly higher than the other two models by 2.2%, indicating its good performance in moderately complex tasks. At the 220 kV voltage level, the accuracy of the research model far exceeds that of other models, especially compared to the MTL model, with an accuracy improvement of 4.6%. This indicates that it has significant advantages in processing high-complexity data. Overall, the research model has shown shorter training and testing times, as well as higher accuracy on datasets of different voltage levels, especially on high complexity 220 kV datasets, where its performance advantage is more pronounced. Two benchmark models of the auto-regression model and back propagation neural network (BPNN) are selected to verify the proposed LSTM–SVM model. Three different data sets are selected for the three load data sets provided for the official websites of three foreign regions. The first is the power load data of Region A from February 1, 2023, to August 31, 2023. The sampling rate of the data is 1 h, and the number of samples is 5,112. The second is the power load data of Region B in 2023, and the sampling rate of the data is 1 h and 8,760 samples. The third is the power load data of Region C in the whole year of 2023, the sampling rate of the data is 1 h, and the sample number is 8,760 samples. The power load data of the three regions are processed by data preprocessing, and the data processed is divided into the dataset. The results of daily load prediction of each model in three different datasets are shown in Figure 9.

Figure 9

Daily load prediction results for each model in three different datasets: (a) Region A, (b) Region B, and (c) Region C.

From Figure 9(a), in Region A, the prediction effect of each model is optimal, and the load prediction curve in the figure almost coincides with the actual situation. In Figure 9(b) and (c), with the increase of mean absolute percent error (MAPE), the prediction value and actual error of each model also increase, reducing the accuracy of the prediction results. Among all the models, only the predicted value of the LSTM–SVM model is the closest to the real load curve, especially at the inflection point. While other models may be able to describe the overall trend of the load curve, their predictive accuracy will diminish as the difficulty of the prediction increases. This can result in significant practical limitations. The MAPE for individual models within 7 days is shown in Figure 10.

Figure 10

Comparison of weekly load prediction errors for each model.

Figure 10 shows that the weekly load prediction accuracy of the LSTM–SVM model is much better than BPNN and AR, while BPNN is higher than AR. A comparison between the predicted load value and the actual load value indicates that the load value prediction reaches an optimal level based on the LSTM–SVM model. This study further simulates the practical application of each model, taking the electricity data of a certain enterprise on a certain day as an example, and dividing it into hourly intervals. Its 24-h electricity load and real-time electricity price are shown in Figure 11.

Figure 11

Real-time change curve of power load and electricity price: (a) 24 h real-time power load and (b) 24 h real-time electricity prices.

In Figure 11(a) and (b), there is a positive correlation between electricity price and electricity load. During the early morning hours, from 1 to 6 o’clock, most users are resting, resulting in a relatively low overall electricity load and corresponding lower electricity prices. As the daytime time increases, the electricity load increases and the electricity price reaches its peak at around 14 o’clock. With the arrival of night, both the electricity load and price show a slight increase, and after 22:00, the price gradually decreases again. This also provides a basis for enterprise electricity planning, which can avoid using too much electricity during peak periods. This study analyzes the performance of the design model in practical applications based on the above data. The extensive use of electricity can cause greenhouse gas emissions and environmental pollution. Therefore, this study takes CO₂ emissions as one of the indicators for emission reduction. In addition, power control is also beneficial for reducing the operating costs of enterprises and can be used as an evaluation indicator. Due to the large amount of data, this study only selects the peak period from 10 am to 5 pm for comparison. Table 5 shows the specific results.

In Table 5, during peak electricity usage, the optimized CO₂ emissions are lower than before optimization. For example, at 10 a.m., the emissions before and after optimization decrease by 15.72%. As the load increases, the gap between the two models also gradually widens. This reduction trend remains consistent throughout the entire time frame, demonstrating the effectiveness of optimization measures in reducing carbon emissions. The optimized cost is lower than before at every time point. For example, at 14:00 in the afternoon, the optimized cost is relatively reduced by 60.48%. In summary, the EDEW system based on LSTM–SVM significantly reduces CO₂ emissions and costs. It can not only help companies reduce emissions but also lower their operating economic costs. To further verify the application effect of the proposed method, the actual operation results of the equipment of the central air conditioning water system in the enterprise are also collected separately, as shown in Table 6.

In Table 6, the difference between the load forecasting results at each time is within ±10%, the predicted total load of the whole day is 28006.7 kW, the actual total load is 27824.9 kW, and the relative error is −0.70%, with a good load forecasting effect. From the perspective of the operation parameters of the equipment, the outlet temperature of the frozen water on the day is set at 11–12°C. One cooling water pump and a cooling tower are in operation, and two chilled water pumps are running. This can assist the enterprise in reducing emissions and the economic cost associated with its operations.

To verify the generality of the proposed method, the research considers the ocean carrier of a multi-energy electric propulsion ship and optimizes the power of its oil generator, energy storage device, and propulsion motor. The comparison results before and after optimization are shown in Table 7.

In Figure 7, the total operating cost of optimized ocean carriers is 359945.42 S, which reduces the total operating cost by 2.1254% compared with that before optimization. The optimized fuel cost of the ocean carriers is 62186.31, a reduction of 1.6320% compared with before the optimization. The optimized ocean carrier reduces a total of 143504.5981 kg of contaminated gas than before the optimization. The results demonstrate that the research method has substantial applicability and universality and offers valuable insights that can be referenced in other industries.