Semiconductor enterprise energy consumption prediction method based on ICEEMDAN

Abstract

The invention relates to a semiconductor enterprise energy consumption prediction method based on ICEEMDAN, which can solve the difficulty that complex dynamic characteristics of energy consumption fluctuation are difficult to capture in nonlinear and non-stationary data processing of a traditional electric power energy consumption prediction method. An improved empirical mode decomposition algorithm and a variational mode decomposition algorithm are adopted to decompose the power consumption. Multi-dimensional factors influencing power consumption are determined through investigation, main influence factors are screened out through improved grey correlation analysis, and finally an optimization prediction algorithm combining a signal decomposition algorithm and a long-short term memory (LSTM) network is provided. According to the algorithm, LSTM model parameters are automatically optimized through an Optuna framework, so that the prediction performance is improved. The method not only theoretically shows the combination potential of a data-driven model and a traditional strategy, but also provides a new energy management tool for a semiconductor enterprise, supports the realization of energy conservation and emission reduction targets of the enterprise, and promotes the stable operation and sustainable development of a power grid.

Description

Technical Field

[0001] The present invention relates to an energy consumption prediction method based on ICEEMDAN, specifically to an energy consumption prediction method for semiconductor enterprises based on ICEEMDAN, which belongs to the field of energy management technology and mainly involves the technical sub-fields of big data analysis, smart buildings and energy prediction. Background Art

[0002] With the intensification of global climate change, energy conservation and emission reduction have become common goals for governments, businesses, and all sectors of society. As the core of modern industry, the semiconductor industry shoulders the heavy responsibility of promoting the development of the digital economy. Its production process consumes a huge amount of energy, especially in processes such as wafer manufacturing, etching, and photolithography, which have very high demands for electricity, pure water, and specialty gases. Unlike traditional manufacturing, the semiconductor industry must not only ensure high precision and a dust-free environment during the production process, but also optimize energy efficiency on this basis. To more effectively manage power resources, control costs, ensure the normal operation of equipment, improve energy efficiency, and respond to environmental protection requirements, developing an accurate power consumption prediction method has become an urgent issue that needs to be addressed.

[0003] Existing power consumption forecasting methods often rely on traditional statistics, linear regression analysis, or time series analysis techniques. These methods often struggle to capture the complex dynamic characteristics of energy consumption fluctuations when dealing with nonlinear and non-stationary power data. Furthermore, existing technologies fail to fully account for the multi-dimensional factors that influence a company's power consumption, such as seasonal variations, the impact of indoor and outdoor temperature and humidity, product type, and company output, limiting the accuracy and generalization capabilities of the forecasting models. Existing forecasting models often perform poorly, especially when responding to emergencies or sudden changes in energy consumption over a short period of time, making it difficult to provide accurate decision support for semiconductor companies' energy management. Summary of the Invention

[0004] The present invention proposes a semiconductor enterprise energy consumption prediction method based on ICEEMDAN, which aims to overcome the above-mentioned shortcomings of the existing technology and improve the accuracy and reliability of enterprise power consumption prediction.

[0005] The technical solution of the present invention is: a semiconductor enterprise energy consumption prediction method based on ICEEMDAN, comprising: using the ICEEMDAN algorithm and the VMD algorithm to preprocess raw power consumption data, separating the main trends and periodic fluctuations in the data, and removing noise and nonlinear components; identifying and integrating key factors affecting the power consumption of semiconductor enterprises, quantifying the correlation between the key factors and the enterprise power consumption from the two dimensions of morphology and distance through an improved grey correlation analysis method, and selecting the main factors closely related to energy consumption as input variables of the model; inputting the processed power consumption data and the screened influencing factors into an optimized LSTM network; and optimizing the key parameters of the LSTM model through the Optuna framework to ensure the optimal configuration of the model.

[0006] Preferably, the method specifically includes the following steps:

[0007] 1) Collect original power consumption data;

[0008] 2) Data preprocessing: This includes verification and cleaning of power consumption data, as well as unified format conversion of seasonal, indoor and outdoor temperature and humidity data, product type, and production data;

[0009] 3) Signal decomposition: The ICEEMDAN algorithm is used to decompose the power consumption data to generate IMFs, and the K-means clustering algorithm is used to classify the IMFs to form comprehensive IMFs, namely Co-IMFs;

[0010] 4) Correlation analysis: Analyze the factors affecting the power consumption of semiconductor enterprises, and use the improved grey correlation analysis method to calculate the correlation between each factor and the power consumption of the enterprise, and select the factors with the highest correlation;

[0011] 5) Input data set construction: The screened influencing factors are used as input and form an input set together with Co-IMFs;

[0012] 6) Power consumption prediction: Use the LSTM model for prediction and Optuna for hyperparameter optimization.

[0013] Preferably, the step 1) includes automatically collecting and calculating the electricity consumption data of each smart meter in real time through a data acquisition system; manually filling in the meter data of non-smart meters in the acquisition system at regular intervals, and the system automatically calculates the electricity consumption of the meters; and estimating the electricity consumption of areas or equipment not covered by the meters based on historical data and experience.

[0014] Preferably, the parameters of the ICEEMDAN algorithm in step 3) include a white noise order of 10 to ensure the quality and reliability of the decomposition.

[0015] Preferably, the factors affecting the power consumption of semiconductor enterprises in step 4) include: 1) season: the season of the data statistics date; 2) indoor and outdoor temperature and humidity: indoor and outdoor temperature and humidity data within the statistical time range; 2) product: product category and output.

[0016] Preferably, the improved grey correlation analysis method in step 4) includes calculating the correlation degree from the distance and shape, and screening out factors with a larger correlation degree.

[0017] Preferably, in step 5), the screened influencing factors are combined with Co-IMF0, Co-IMF1 and Co-IMF2 to form three input sets.

[0018] Preferably, the hyperparameter space of the LSTM model in step 6) preferably includes the optimizer, the number of units in the LSTM layer, the dropout rate, the batch size and the learning rate, and the optimal parameter combination is automatically found through the Optuna library.

[0019] Preferably, the LSTM structure in step 6) includes an input layer, LSTM layer 1+Dropout layer 1, LSTM layer 2+Dropout layer 2, LSTM layer 3+Dropout layer 3, LSTM layer 4+Dropout layer 4 and a Dense layer.

[0020] The advantages of this invention include a rationally designed method that utilizes advanced ICEEMDAN signal processing technology to decompose complex energy consumption data into easily analyzed and predicted sub-signals, effectively extracting key features from the data, and accurately predicting future energy consumption trends using deep learning algorithms. This method not only theoretically demonstrates the potential of combining data-driven models with traditional energy management strategies, but also provides a new energy management tool for semiconductor companies in practice, enabling them to achieve energy conservation and emission reduction goals while also supporting the stable operation and sustainable development of the power grid. It can more accurately process and analyze the nonlinear characteristics of enterprise power consumption data, thereby improving the accuracy and reliability of enterprise power consumption forecasts. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of the semiconductor enterprise energy consumption prediction method based on ICEEMDAN of the present invention.

[0022] Figure 2 It is a schematic diagram of the hybrid prediction algorithm model.

[0023] Figure 3 This is a schematic diagram of the LSTM neural network structure.

[0024] Figure 4-1 FIG. 4 is a diagram showing the ICEEMDAN decomposition results of the ICEEMDAN decomposition analysis according to an embodiment of the present invention.

[0025] Figure 4-2 4 is a correlation heat map of the ICEEMDAN decomposition analysis according to an embodiment of the present invention.

[0026] Figure 5-1 4 is a sample entropy calculation result diagram of the sample entropy result analysis according to an embodiment of the present invention.

[0027] Figure 5-2 2 is a box plot of the sample entropy analysis results of the embodiment of the present invention.

[0028] Figure 6 2 is a cluster analysis result diagram of an embodiment of the present invention.

[0029] Figure 7 is a Co-IMFs diagram after integration according to an embodiment of the present invention.

[0030] Figure 8-1 4 is a diagram of VMD decomposition results of VMD decomposition analysis according to an embodiment of the present invention.

[0031] Figure 8-2 4 is a correlation heat map of the VMD decomposition analysis according to an embodiment of the present invention.

[0032] Figure 9-1 This is a graph showing changes in the objective function value of the Optuna-LSTM algorithm during the optimization process of the Optuna-LSTM algorithm according to an embodiment of the present invention.

[0033] Figure 9-2 This is a diagram illustrating the importance of Optuna-LSTM algorithm parameters during the optimization process of the Optuna-LSTM algorithm according to an embodiment of the present invention.

[0034] Figure 10 This is a graph showing the change in the Optuna-LSTM loss function according to an embodiment of the present invention.

[0035] Figure 11 This is a diagram of the Optuna-LSTM prediction results in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described in detail below with reference to examples and specific implementation methods.

[0037] To solve the above technical problems, the present invention proposes a semiconductor enterprise energy consumption prediction method based on ICEEMDAN. First, the improved empirical mode decomposition (ICEEMDAN) and variational mode decomposition (VMD) algorithms are used to preprocess the original power energy consumption data, effectively separating the main trends and periodic fluctuations in the data, while removing noise and nonlinear components, improving the clarity and analyzability of the data, and providing more accurate input for subsequent deep learning prediction models.

[0038] Based on historical energy management project experience and field research at semiconductor companies, we identified and integrated key factors influencing semiconductor companies' power consumption, including seasonality, indoor and outdoor temperature and humidity, product type, and production volume. Using an improved grey correlation analysis method, we quantified the correlation between these factors and companies' power consumption from the perspectives of both form and distance. This allowed us to select the key factors most closely related to energy consumption as input variables for the model.

[0039] The processed power consumption data and the identified influencing factors were fed into the optimized LSTM network. The LSTM network consists of multiple layers, including an input layer, multiple hidden layers, and an output layer. Each hidden layer is configured with a dropout layer to prevent overfitting. Key parameters of the LSTM model (such as the number of hidden layer units, batch size, and learning rate) were optimized using the Optuna framework to ensure optimal model configuration.

[0040] Specifically, such as Figure 1 As shown, a semiconductor enterprise energy consumption prediction method based on ICEEMDAN includes the following steps:

[0041] 1) Data collection through smart meters: Smart meters are installed in necessary equipment and areas to collect enterprise power consumption data, and are connected to the data collection system through wired or wireless means to realize intelligent collection of smart meter data.

[0042] 2) Data preprocessing: This includes the verification and cleaning of power consumption data, as well as the unified format conversion of meteorological data (season, indoor and outdoor temperature and humidity data), product types, and output data to ensure the validity and consistency of relevant data.

[0043] The processing of product types and output data is preferably as follows: using date as the main line to count the daily electricity consumption of the enterprise, the types of products produced and the corresponding product output.

[0044] The preferred method for acquiring power consumption data is: automatically collecting and calculating the power consumption data of each smart meter in real time through a data acquisition system; manually filling in the meter data of non-smart meters in the acquisition system on time, and the system automatically calculates the power consumption of the meter; estimating the power consumption of areas or equipment that are not covered by the meters (including smart meters and ordinary meters) through historical data reference and experience, so as to supplement the power consumption of areas not covered by the meters, so that the statistical total power consumption is as close to the actual power consumption as possible.

[0045] 3) Signal decomposition: The ICEEMDAN algorithm is used to decompose the power consumption data to generate IMFs, and the K-means clustering algorithm is used to classify the IMFs to form comprehensive IMFs (Co-IMFs).

[0046] The parameters of the ICEEMDAN algorithm preferably include a white noise number of 10 to ensure the quality and reliability of the decomposition.

[0047] 4) Correlation Analysis: Based on historical experience and on-site survey analysis, the factors affecting the power consumption of semiconductor companies are obtained. The improved grey correlation analysis method is used to calculate the correlation between each influencing factor and the power consumption of the company, and the influencing factors with the highest correlation are screened out.

[0048] The key influencing factors obtained from the investigation and analysis include: 1) Season: the season of the data statistics date; 2) Indoor and outdoor temperature and humidity: indoor and outdoor temperature and humidity data within the statistical time range; 2) Product: product category and output, etc.

[0049] The improved grey correlation analysis method preferably includes calculating the correlation degree from the distance and shape, screening out the factors with greater correlation degree, among which the geometric shape is more important than the distance, and the trend and shape can better reflect the potential relationship between different factors, so ω1=0.7, ω2=0.3; the consistency coefficient is ξ=0.7.

[0050] 5) Input data set construction: The screened influencing factors are used as input and form three input sets together with Co-IMFs.

[0051] 6) Power consumption prediction: Use the LSTM model for prediction and Optuna for hyperparameter optimization.

[0052] The LSTM model's hyperparameter space preferably includes the optimizer, the number of LSTM layer units, the dropout rate, the batch size, and the learning rate. The Optuna library automatically finds the optimal parameter combination. The LSTM structure preferably includes the input layer, LSTM layer 1 + Dropout layer 1, LSTM layer 2 + Dropout layer 2, LSTM layer 3 + Dropout layer 3, LSTM layer 4 + Dropout layer 4, and a Dense layer.

[0053] In order to further specifically illustrate the technical solution of the present invention, the signal decomposition algorithm, correlation analysis and prediction algorithm involved above are introduced in detail.

[0054] 1) Signal decomposition algorithm

[0055] ICEEMDAN is an improved signal processing method specifically designed for analyzing nonlinear and nonstationary time series data. Based on the principles of empirical mode decomposition (EMD), this method improves upon the classic EMD method by introducing adaptive noise into the original signal. In ICEEMDAN, the signal is decomposed multiple times, each time with varying amounts of white noise added. The resulting intrinsic mode functions (IMFs) are then averaged. ICEEMDAN effectively reduces modal aliasing and boundary effects, improving the stability and reliability of the signal decomposition. Compared to traditional EMD, ICEEMDAN can more accurately extract key features from the signal.

[0056] The variational mode decomposition (VMD) algorithm is an efficient signal processing method used to decompose complex signals into multiple intrinsic mode functions. VMD extracts each mode of the signal by establishing a constrained variational problem. At the same time, the fidelity and robustness of the reconstructed signal are maintained through the alternating direction multiplication method in the iterative process. Compared with other signal decomposition techniques such as ICEEMDAN, the main advantage of VMD is that it can more stably and effectively handle nonlinear and non-stationary characteristics in the data, and has better robustness to various types of data noise. Therefore, when processing complex power consumption data, VMD can accurately identify and extract key patterns and trends in power consumption, thereby providing a more reliable foundation for subsequent analysis and prediction.

[0057] Further applying the VMD algorithm to the IMFs with distinct high-frequency characteristics obtained after decomposition using the ICEEMDAN algorithm enables more detailed and accurate analysis of these high-frequency signals, enabling deeper extraction of key information from power consumption data. Specifically, ICEEMDAN effectively extracts components of different frequencies from the raw data, and applying the VMD algorithm to these high-frequency IMFs further extracts more detailed frequency components and features, enabling a better understanding and analysis of the dynamic changes in power consumption, resulting in a more refined and accurate signal representation, providing richer and more accurate input data for subsequent power consumption analysis and forecasting.

[0058] 2) Correlation analysis

[0059] Deng's grey relational model fully considers nonlinear and heterogeneous factors and introduces a weight adjustment mechanism, allowing for more flexible weighting of factors, thereby improving the accuracy and reliability of the correlation. The model determines the correlation by calculating the similarity of the geometric shape and change trend between the factor sequence and the comparison sequence (Formula (1)).

[0060] The improved GRA method adds sequence distance analysis on the basis of the Deng model, combines the distance parameter and the grey correlation degree, comprehensively considers the curve distance and shape, and evaluates the consistency between sequences through the consistency correlation degree, making the analysis results more comprehensive and accurate.

[0061]

[0062] Wherein, X0={x0(1),x0(2),…,x0(n)} is a factor sequence, which in the present invention mainly refers to an electric power sequence, and x0(k) is the observed value of X0 at the kth point; i ={x i (1),x i (2),…,x i (n)} is the comparison sequence, which mainly refers to the relevant factors affecting power consumption in the present invention; ρ is the resolution coefficient, and the value range is generally ρ∈(0,1). The smaller ρ is, the better the resolution is, and ρ=0.5 is usually taken; γ represents the degree of similarity between the factor sequence and the comparison sequence. The larger γ is, the stronger the correlation between the two and the closer the relationship is.

[0063]

[0064] in, It is X0 and X i The average distance between them, that is, the relative position of the two curves; Is the distance parameter, which is used to unify the order of magnitude of each sequence, thereby eliminating the distance between the sequence values of the factor sequence. The impact of δ D ≥0, and δ D The smaller the value, the closer the relative positions of the two sequences are. is the mean of the factor series, is the distance from the factor sequence to the x-axis.

[0065] γ C =ω1(1-δ Di )+ω2γ (3)

[0066] Among them, ω1 and ω2 are about the distance parameter δ D The relative weight with the grey correlation degree γ satisfies ω1+ω2=1. ω1 and ω2 are selected according to the specific application background. In most sequence application scenarios, distance and shape are generally considered to be equally important, that is, ω1=ω2=0.5. The calculated consistency correlation degree γ C It needs to be compared with the consistency coefficient ξ specified in advance. If γ C If the value of ξ is greater than or equal to ξ, the two sequences are considered to be consistent, otherwise they are considered inconsistent. The value of ξ should be determined based on the characteristics of the reference sequence and the specific application.

[0067] 3) Prediction Algorithm

[0068] The Optuna-LSTM algorithm combines Optuna's automatic hyperparameter optimization capabilities with the strengths of LSTM networks in processing time series data, providing a powerful and efficient forecasting tool. This algorithm automatically searches for optimal hyperparameter configurations, such as the number of network layers, number of hidden units, and learning rate, through the Optuna optimization framework. This significantly reduces the workload of manual parameter tuning and the uncertainty associated with relying on empirical experience. Furthermore, the LSTM network's ability to capture long-term dependencies makes the model particularly suitable for forecasting complex time series data, effectively improving forecast accuracy and stability.

[0069] The flexibility and ease of implementation of the Optuna-LSTM algorithm enable it to be easily integrated into existing machine learning libraries and frameworks, adapting to a variety of different forecasting scenarios, thereby providing an efficient, accurate, and widely applicable solution for time series forecasting.

[0070] like Figure 2 As shown, a semiconductor enterprise energy consumption prediction method based on ICEEMDAN further specifically includes the following steps:

[0071] 1) Data preprocessing

[0072] The sources of electricity consumption data include: real-time automatic collection and calculation of electricity consumption data of each smart meter through the data acquisition system; manual reporting of meter data of non-smart meters in the acquisition system at regular intervals, and automatic calculation of meter electricity consumption by the system; and electricity consumption of areas or equipment not covered by meters (including smart meters and ordinary meters) estimated based on historical data and experience.

[0073] Due to the diverse data sources and the potential for errors during data transmission or entry, the accuracy of the acquired data must be verified, with values significantly deviating from the normal range flagged and reviewed. Data marked as abnormal undergoes data cleaning and is compared with historical data or data from adjacent time points to determine whether it needs to be corrected or deleted. To facilitate subsequent data analysis, the system automatically standardizes the data format, including, but not limited to, converting timestamps to a unified date and time format and converting power consumption data to consistent units.

[0074] Temperature and humidity data are primarily derived from monitoring data from relevant temperature and humidity sensors deployed by the enterprise. Key parameters include indoor and outdoor temperature and humidity. To address potential data loss during sensor monitoring, this invention uses a mean-filling method to automatically fill in missing data. This method calculates the mean of the monitoring data from three adjacent monitoring time points before and after the missing data, thereby ensuring the continuity and integrity of the monitoring data. To ensure the validity and consistency of temperature and humidity data in subsequent analysis, all temperature and humidity data are converted into a unified unit of measurement and data format to ensure data uniformity.

[0075] Product output data is based on the data from the enterprise's production management system. Data sources include manual reporting and direct automatic acquisition of daily output of each category of products through the enterprise's MES system connection.

[0076] 2) Signal decomposition

[0077] The specific steps for decomposing the power consumption after preprocessing are as follows:

[0078] First, the ICEEMDAN algorithm was used to decompose the power consumption data, generating multiple IMFs, each representing a specific frequency component of the data. The number of white noise additions was set to 10 to ensure the quality and reliability of the decomposition. EMD was performed after each addition of white noise, and all resulting IMFs were averaged to reduce modal aliasing and boundary effects, thereby improving the stability of the decomposition.

[0079] Next, we calculate sample entropy for each IMF. Sample entropy measures the complexity and regularity of a data sequence. Higher sample entropy indicates a more complex and irregular data sequence, helping to identify important features and patterns in the data and providing a basis for subsequent clustering and integration. This stage uses the functionality provided by the SampEn library to calculate sample entropy. During the calculation, we set the embedding dimension (mm) to 1 to capture the complex dynamics of the data, and the tolerance (r) to 0.1 to more accurately assess the similarity between data sequences.

[0080] Next, the K-means clustering algorithm is used to classify IMFs based on sample entropy, which helps identify IMFs with similar characteristics. This step uses the KMeans function in the sklearn.cluster module to implement the clustering operation. By setting the number of clusters (num_clusters) to 3, the IMFs are divided into three different groups, each containing IMFs with similar sample entropy characteristics. In addition, the random_state parameter is set to 9 to ensure consistent results between code runs. The n_init parameter during the clustering process is set to 'auto', allowing the algorithm to automatically select the initial value of the cluster center to improve classification accuracy.

[0081] Next, the IMFs belonging to the same category obtained by K-means clustering are combined into several composite IMFs (Co-IMFs). For example, all IMFs belonging to the first category are added together to form the first Co-IMF, and so on. The goal is to reduce the complexity of the data while retaining key information.

[0082] Finally, in the final stage of power consumption data processing, the clustered Co-IMF0, which has prominent high-frequency features, is subjected to a secondary decomposition using the VMD method to further refine and extract key patterns and trends in the power consumption data. This process uses the VMD function in the vmdpy library to implement VMD decomposition. The VMD algorithm controls the decomposition process by setting a series of parameters such as alpha, tau, K, DC, init, and tol.

[0083] 3) Correlation analysis

[0084] The selected influencing factors are subjected to correlation analysis. The improved grey correlation analysis method is used to calculate the correlation between the selected influencing factors and power consumption from the perspectives of distance and shape. The specific method is shown in formula (3). The factors with a greater correlation with power consumption are selected as part of the input of the prediction model.

[0085] 4) Input data set construction

[0086] The screened influencing factors are combined with Co-IMF0, Co-IMF1 and Co-IMF2 to form three input sets in preparation for subsequent predictions.

[0087] 5) Power consumption forecast

[0088] First, the data set is divided: use the train_test_split function in the sklearn.model_selection library to divide the input set into training set, validation set and test set in a ratio of 6:3:1.

[0089] Secondly, data preprocessing: Use the MinMaxScaler function in the sklearn.preprocessing library to normalize the data and convert the features to the range of 0 to 1; use the TimeseriesGenerator module in tensorflow.keras.preprocessing.sequence to create a time series dataset, create generators for the training set, validation set, and test set respectively, and extract features and labels for subsequent model training, validation, and testing.

[0090] Then, LSTM hyperparameter optimization: define an optimization objective function, which includes a hyperparameter space, including the optimizer, the number of units in the LSTM layer, the dropout rate, the batch size, the learning rate, etc.; use Sequential to build a four-layer LSTM model (such as Figure 3 The model uses mean square error as the loss function and is compiled based on the optimizer and learning rate in the hyperparameter space. The Optuna library is used to automatically find the optimal combination of model parameters to achieve the most accurate analysis of power consumption data.

[0091] Finally, LSTM model prediction: Optuna automatically finds the optimal parameters for model training, and the trained model is used to make predictions on the test set to verify the model's prediction effect.

[0092] Example: A large semiconductor manufacturer in Wuxi

[0093] Through research and analysis of historical data, we found that the main factors affecting an enterprise's electricity consumption are: indoor and outdoor temperature and humidity, season, product type and output, etc.

[0094] The ICEEMDAN-based semiconductor enterprise energy consumption prediction method includes the following steps:

[0095] 1) Preprocess the acquired relevant power data, process the missing values and outliers in the power consumption data respectively, and process the collected product output and type data with daily time granularity.

[0096] 2) Use the ICEEMDAN algorithm to decompose the daily power consumption data for the first time and calculate the correlation between each IMF. The decomposition results are analyzed as follows: Figure 4-1 、 4-2 The results show that each IMF component fluctuates mainly around x = 0, showing the oscillation characteristics of the signal, and each IMF component is independent of each other. It can be considered that the decomposition effect of ICEEMDAN is good.

[0097] The sample entropy of the decomposition results was calculated, and different embedding dimensions and tolerances were selected to evaluate the system's complexity. The results are shown in Figure 5. The mean sample entropy calculated with the parameter settings of m = 2 and r = 0.1 is 0.60795, indicating that the corresponding IMF components can capture a high degree of complexity under these parameters. The standard deviation is 0.78605, indicating that the sample entropy values show large fluctuations between different IMF components. Therefore, the sample entropy under this parameter configuration can reveal the complexity and volatility of the signal.

[0098] Clustering the sample entropy results, the results are as follows Figure 6As shown in Figure 2, the sample entropy calculation results are divided into three different categories. IMF0 and IMF1 are classified into the same category because the sample entropy calculation results of these two components are similar, that is, their time series signals have high complexity and irregularity; IM2, IMF3 and IMF4 are classified into one category because the complexity of their signals is between category 1 and category 0; IMF5 to IMF11 are all classified into the third category. The integrated Co-IMFs are shown in Figure 2. Figure 7 As shown. Since Co-IMF0 has obvious high-frequency features, VMD decomposition is applied to it, and the results are as follows Figure 8-1 、 8-2 shown.

[0099] 3) For the influencing factors proposed in the survey, the total power is used as the factor sequence, and the 11 influencing factors such as indoor and outdoor temperature and humidity, product type and corresponding output are used as the comparison sequence. The grey correlation degree γ and distance parameter δ are calculated according to formula (1) and formula (2) respectively. D , where the resolution coefficient ρ = 0.5. Finally, the calculated γ and δ D , put it into formula (3), and get the final consistency correlation γ C Combined with the application background of the enterprise power energy consumption of the present invention, ω1=0.7, ω2=0.3; the consistency coefficient is ξ=0.7, and the calculation results are shown in Table 1.

[0100] Table 1 Improved grey relational calculation results

[0101]

[0102]

[0103] 4) The influencing factors obtained by screening are combined with the Co-IMFs obtained by the signal decomposition algorithm to form three data sets containing influencing factors, preparing for the next step of prediction.

[0104] 5) Define the hyperparameter search space of Optuna. The specific definition is shown in Table 2. After Optuna's automatic search function, the optimal parameters of the experiment are quickly located. See Table 2. The specific optimization process of the Optuna-LSTM algorithm is as follows: Figure 9-1 、 9-2 As shown in the figure, the algorithm uses RMSE as the objective function value. As the number of experiments increases, the objective value decreases rapidly at the beginning, indicating that the model performance is rapidly improved by adjusting the parameters in the early stage of optimization. As the number of experiments increases, the rate of decrease of the objective function value slows down and gradually stabilizes at the fifth experiment, indicating that the parameter settings are gradually approaching the optimal solution. Finally, the RMSE stabilizes at around 25. After the optimization is completed, the importance of each parameter is analyzed as follows. Figure 9-2 shown.

[0105] Table 2 LSTM parameter setting table

[0106]

[0107]

[0108] 6) Use the optimal parameters obtained by screening to train, verify and predict the LSTM deep learning model. The experiment selected MSE as the loss function and iterated 200 times to obtain the final result. The loss function of the algorithm changes during training as shown below: Figure 10 As shown in the figure, by observing the rate of decrease of the loss function of each algorithm to evaluate the convergence speed, it can be seen that Optuna converges fastest in the initial iterations, and shows a downward trend as the number of iterations increases. The downward curve is the most stable, indicating good stability, and the average calculation speed is about 1.001.

[0109] The final prediction results are as follows Figure 11 The evaluation index scores were calculated based on the prediction results, as shown in Table 3. The Optuna-LSTM algorithm fits the actual electric power curve well, with an RMSE of 24.41, indicating a small average deviation between the predicted and actual values. The MAE was 4.01, indicating that the average absolute error between the predicted and actual values is within an acceptable range. The MAPE was 3.07%, indicating a low percentage of the average error of the predicted values to the actual values. The R² was 0.89, indicating a high correlation between the model's predictions and the actual values.

[0110] Table 3 Evaluation index scores

[0111]

[0112] The above description is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this field, several variations and improvements can be made without departing from the creative concept of the present invention, which all fall within the scope of protection of the present invention.

Claims

1. A semiconductor enterprise energy consumption prediction method based on ICEEMDAN, characterized in that: This includes preprocessing the raw power consumption data using the ICEEMDAN algorithm and VMD algorithm to separate the main trends and periodic fluctuations in the data while removing noise and nonlinear components; Identify and integrate the key factors affecting the power energy consumption of semiconductor companies. Using an improved grey correlation analysis method, quantify the correlation between these key factors and the company's power energy consumption from the two dimensions of morphology and distance, and select the main factors closely related to energy consumption as the input variables of the model. Input the processed power consumption data and the screened influencing factors into the optimized LSTM network. Optimize the key parameters of the LSTM model using the Optuna framework to ensure the optimal configuration of the model.

2. The semiconductor enterprise energy consumption prediction method based on ICEEMDAN according to claim 1, characterized in that: The specific steps include: 1) Collect original power consumption data; 2) Data preprocessing: This includes verification and cleaning of power consumption data, as well as unified format conversion of seasonal, indoor and outdoor temperature and humidity data, product type, and production data; 3) Signal decomposition: The ICEEMDAN algorithm is used to decompose the power consumption data to generate IMFs, and the K-means clustering algorithm is used to classify the IMFs to form comprehensive IMFs, namely Co-IMFs; 4) Correlation analysis: Analyze the factors affecting the power consumption of semiconductor enterprises, and use the improved grey correlation analysis method to calculate the correlation between each factor and the power consumption of the enterprise, and select the factors with the highest correlation; 5) Input data set construction: The screened influencing factors are used as input and form an input set together with Co-IMFs; 6) Power consumption prediction: Use the LSTM model for prediction and Optuna for hyperparameter optimization.

3. The semiconductor enterprise energy consumption prediction method based on ICEEMDAN according to claim 2, characterized in that: The step 1) includes automatically collecting and calculating the electricity consumption data of each smart meter in real time through the data acquisition system; manually filling in the meter data of non-smart meters in the acquisition system at regular intervals, and the system automatically calculates the electricity consumption of the meters; and estimating the electricity consumption of areas or equipment not covered by the meters based on historical data and experience.

4. The semiconductor enterprise energy consumption prediction method based on ICEEMDAN according to claim 2, characterized in that: The parameters of the ICEEMDAN algorithm in step 3) include the number of white noises being 10.

5. The semiconductor enterprise energy consumption prediction method based on ICEEMDAN according to claim 2, characterized in that: The factors affecting the power consumption of semiconductor enterprises in step 4) include: 1) season: the season of the data statistics date; 2) indoor and outdoor temperature and humidity: indoor and outdoor temperature and humidity data within the statistical time range; 2) product: product category and output.

6. The semiconductor enterprise energy consumption prediction method based on ICEEMDAN according to claim 2, characterized in that: The improved grey correlation analysis method in step 4) includes calculating the correlation degree from the distance and shape, and screening out factors with a larger correlation degree.

7. The semiconductor enterprise energy consumption prediction method based on ICEEMDAN according to claim 2, characterized in that: In step 5), the influencing factors obtained by screening are combined with Co-IMF0, Co-IMF1 and Co-IMF2 to form three input sets.

8. The method for predicting energy consumption of semiconductor enterprises based on ICEEMDAN according to claim 2, characterized in that: The hyperparameter space of the LSTM model in step 6) preferably includes the optimizer, the number of units in the LSTM layer, the dropout rate, the batch size, and the learning rate, and the optimal parameter combination is automatically found through the Optuna library.

9. The semiconductor enterprise energy consumption prediction method based on ICEEMDAN according to claim 8, characterized in that: The LSTM structure in step 6) includes an input layer, LSTM layer 1+Dropout layer 1, LSTM layer 2+Dropout layer 2, LSTM layer 3+Dropout layer 3, LSTM layer 4+Dropout layer 4 and a Dense layer.

Images