A Semiconductor Factory Energy Data Prediction Method Based on the Fusion of Multiple Time Series Models

Abstract

This invention discloses a method for predicting energy data in semiconductor factories based on the fusion of multiple time-series models. The method includes: cleaning and defining the features of energy data; selecting and training multiple individual models based on the data characteristics of the processed energy data; selecting a suitable individual model based on scenario discrimination indicators, wherein the scenario discrimination indicators include fluctuation amplitude and seasonal cycle; the fluctuation amplitude is calculated by the standard deviation of the data or the range within a moving window, and the seasonal cycle is obtained by detecting the cycle strength using an autocorrelation function; fusing the prediction results of the selected models; evaluating the fused model and making energy predictions based on the evaluation results. This invention can adaptively select the optimal model, ensuring the accuracy and stability of energy prediction.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing technology, and in particular relates to a method for predicting energy data in semiconductor factories based on the fusion of multiple time-series models. Background Technology

[0002] Semiconductor factories are typical manufacturing scenarios characterized by "high energy consumption, high precision, and high stability." The stability of energy supply directly determines wafer yield (for example, power fluctuations may cause abnormal light source intensity in lithography machines, leading to batch defects). Therefore, accurate energy forecasting has become a core requirement for smart factory management in semiconductor factories.

[0003] In semiconductor plant systems, energy data (such as electricity, water, and steam consumption) exhibits strong time-series characteristics, seasonality, nonlinearity, and sudden fluctuations (such as equipment start-up and shutdown). Existing technologies typically employ a single model (such as statistical models ARIMA and SARIMA, or deep learning models LSTM and Transformer) for prediction. While these models perform well in specific scenarios, they cannot simultaneously capture the multiple features of the data. For example, statistical models are suitable for handling linear trends and seasonal data, but their accuracy decreases under nonlinear and complex relationships; deep learning models excel at long-term dependencies but are sensitive to data drift and exhibit poor prediction stability. Therefore, the main technical problem with existing technologies is that a single model cannot adaptively select the optimal model based on the specific scenario characteristics of the data (such as fluctuation amplitude and seasonal cycle), resulting in insufficient accuracy and stability in energy data prediction. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a semiconductor factory energy data prediction method based on the fusion of multiple time series models, which can adaptively select the optimal model and ensure the accuracy and stability of energy prediction, in order to address the shortcomings of the prior art.

[0005] The first aspect of this invention discloses a method for predicting energy data in semiconductor factories based on the fusion of multiple time-series models, comprising the following steps:

[0006] S1: Perform data cleaning and feature definition on energy data. Data cleaning includes removing outliers and filling in missing values. Feature definition includes constructing time-series features, external features, and derived features.

[0007] S2: Based on the data characteristics of the energy data processed in step S1, select a model and train multiple single models. The single model is a statistical model, a deep learning model, or a hybrid model. The training adopts time series cross-validation.

[0008] When selecting a model, choose the type of single model based on the characteristics of the data: statistical models are used for linear trend data; deep learning models are used for non-linear or long-sequence data; and hybrid models are used for complex scenarios.

[0009] S3: Select a suitable single model based on the scene discrimination index, which includes fluctuation amplitude and seasonal cycle; the fluctuation amplitude is calculated by the standard deviation of the data or the range within the moving window, and the seasonal cycle is obtained by detecting the cycle intensity through the autocorrelation function;

[0010] When mapping scenario discrimination indicators to model fusion selection: if the fluctuation range exceeds the threshold, deep learning models are preferred; if the seasonal cycle exceeds the threshold, statistical models are preferred; otherwise, hybrid models are selected.

[0011] S4: Perform model fusion on the prediction results of the selected model;

[0012] S5: Evaluate the fusion model and make energy forecasts based on the evaluation results.

[0013] In the above-mentioned energy data prediction method, in S1, the derived features include time-domain features and frequency-domain features. The time-domain features are calculated by rolling mean or rolling variance, and the frequency-domain features are extracted by Fourier transform or wavelet transform to enhance the model's ability to capture periodicity.

[0014] In the above energy data prediction method, in S1, the external features include production plan quantity, equipment operating status, and environmental parameters, including temperature and humidity.

[0015] In the above energy data prediction method, in S2, the statistical model includes ARIMA or SARIMA, the deep learning model includes LSTM, Seq2Seq or Transformer, and the hybrid model includes a combination of ARIMA-LSTM.

[0016] In the above energy data prediction method, in S3, the scenario discrimination index includes a fluctuation amplitude threshold. When the fluctuation amplitude exceeds the preset threshold, a deep learning model is selected first; when the seasonal cycle characteristics are obvious, a statistical model is selected first.

[0017] In the aforementioned energy data prediction method, S3 selects the optimal model or model combination through a rule engine or classifier. The rule engine is built based on decision trees or logistic regression and is used to map scenario indicators to model selection.

[0018] In the aforementioned energy data prediction method, S4 employs a weighted average method for model fusion. The weights for this weighted average fusion are optimized through regression learning, including using Ridge regression to fit the relationship between each model's predicted values ​​and the true values, and solving for the optimal weights. The optimization formula for Ridge regression is:

[0019] ;

[0020] in, Represents a weight vector, with elements The weights corresponding to the j-th model; Represents the actual value; This represents the weighted average predicted value, calculated using the following formula: ,in Let represent the predicted value of the j-th model for sample i; n represent the number of samples; m represent the number of models; λ represents the regularization parameter used to control the weight complexity; the solution is obtained by cross-validation to determine the optimal λ and . .

[0021] In the above energy data prediction method, in S4, when the model effects are similar, the arithmetic mean or geometric mean is used to fuse the prediction results of the selected model.

[0022] In the above energy data prediction method, S5 uses mean square error or mean absolute error for evaluation, and stores the model when the evaluation result meets the preset accuracy requirements.

[0023] In the energy data prediction method described above, S2 employs a time-series cross-validation method for training. This involves dividing the data into multiple time windows, with each window serving as the validation set for iterative training to avoid overfitting.

[0024] The second aspect of this invention discloses a semiconductor factory energy data prediction system based on the fusion of multiple time-series models, including a data processing module, a model training module, a model selection module, a model fusion module, and a model evaluation module;

[0025] The data processing module is used to perform data cleaning and feature definition on energy data. Data cleaning includes removing outliers and filling in missing values, while feature definition includes constructing time-series features, external features, and derived features.

[0026] The model training module is used to select and train multiple single models based on the data characteristics of the energy data processed by the data processing module. The single model can be a statistical model, a deep learning model, or a hybrid model. The training adopts time series cross-validation.

[0027] The model selection module is used to select a suitable single model based on scenario discrimination indicators, which include fluctuation range and seasonal cycle. Fluctuation range is calculated by the standard deviation of the data or the range within the moving window, and seasonal cycle is obtained by detecting the cycle intensity through the autocorrelation function.

[0028] The model fusion module is used to fuse the prediction results of the selected models.

[0029] The model evaluation module is used to evaluate the fusion model and make energy predictions based on the evaluation results.

[0030] A third aspect of the present invention discloses an electronic device, comprising: a memory and a processor, wherein the processor and the memory are connected;

[0031] The memory is used to store programs;

[0032] The processor invokes a program stored in the memory to execute the method provided by the first aspect embodiment and / or any possible embodiment in combination with the first aspect embodiment.

[0033] The fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a computer to perform the methods provided by the embodiments of the first aspect and / or any possible embodiments in combination with the embodiments of the first aspect.

[0034] Compared with existing technologies, this invention has the following advantages: It solves the problem that a single model cannot adaptively select the optimal model based on data scenario characteristics, thereby improving the accuracy and stability of energy prediction. Specifically, step S3 dynamically selects a suitable single model based on specific scenario discrimination indicators (including fluctuation amplitude and seasonal cycle). The fluctuation amplitude is calculated using the standard deviation of the data or the range within a moving window, and the seasonal cycle is obtained by detecting the cycle strength using the autocorrelation function. This feature allows the method to respond to data changes in real time: when the fluctuation amplitude is large, a deep learning model with strong nonlinear capture ability is prioritized; when the seasonal cycle is obvious, the statistical model is strengthened, overcoming the limitations of a single model in complex scenarios. Combining the model training in step S2 (using time series cross-validation) and the model fusion in step S4, this method reduces prediction errors and improves overall robustness through complementary advantages, maintaining stable output even when data experiences sudden fluctuations or drift. Finally, the evaluation and prediction in step S5 ensure the reliability of the scheme, achieving high accuracy and reliability in energy consumption prediction.

[0035] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0036] Figure 1This is a flowchart of the method in Embodiment 1 of the present invention.

[0037] Figure 2 This is a system module diagram of Embodiment 2 of the present invention. Detailed Implementation

[0038] like Figure 1 As shown, a method for predicting energy data in semiconductor factories based on the fusion of multiple time-series models includes the following steps:

[0039] S1: Perform data cleaning and feature definition on energy data. Data cleaning includes removing outliers and filling in missing values. Feature definition includes constructing time-series features, external features, and derived features.

[0040] In practice: First, data cleaning includes outlier removal and missing value imputation. Outlier removal uses the Z-score method: calculate the Z-score of each data point (i.e., (data value - data mean) / standard deviation). If the absolute value of the Z-score exceeds 3 (the industry standard threshold), it is considered an outlier and removed. Missing value imputation uses linear interpolation: for missing time points (e.g., missing electricity consumption in a certain hour), fill them with the linear trend of adjacent data points (e.g., the average of the previous and next points). Feature definition includes constructing three types of features: time-series features (e.g., rolling mean and rolling variance), external features (e.g., production plan quantities), and derived features (e.g., frequency domain features). Time-series features are calculated using a sliding window (e.g., calculating the rolling mean of the past 24 hours, representing a short-term trend); external features are obtained from relevant systems (e.g., reading production plan quantities from the factory's MES system); derived features are constructed through mathematical transformations.

[0041] In the power consumption forecasting of a semiconductor factory, hourly electricity consumption data is input. Data cleaning includes: detecting outliers (e.g., removing a point with a Z-score of 4), and imputing missing values ​​(e.g., linear interpolating 11:00 and 13:00 data for missing 12:00 data). Feature definition includes: calculating the rolling mean (average electricity consumption over the past 24 hours), adding external features (daily production plan), and deriving features. The output is the cleaned dataset and feature vectors for subsequent steps.

[0042] Data cleaning improves data quality because outliers (such as sensor malfunctions) can lead to incorrect model learning; filling in missing values ​​ensures temporal continuity and avoids prediction bias. Feature definition enhances model input information; temporal features capture trend changes, external features introduce context (such as production activities), and derived features strengthen the capture of periodicity.

[0043] S2: Based on the data characteristics of the energy data processed in step S1, select a model and train multiple single models. The single model is a statistical model, a deep learning model, or a hybrid model. The training adopts time series cross-validation.

[0044] When selecting a model, choose the type of single model based on the characteristics of the data: for example, statistical models are used for linear trend data; deep learning models are used for non-linear or long sequence data; hybrid models are used for complex scenarios; other models can also be selected based on the characteristics of the data.

[0045] Statistical models (such as ARIMA or SARIMA) are used for linear trend data (such as seasonally significant electricity consumption); deep learning models (such as LSTM or Transformer) are used for non-linear, long-sequence data (such as fluctuations caused by equipment start-up and shutdown); hybrid models (such as ARIMA-LSTM) are used for complex scenarios (such as those with both trends and residuals). Training employs time-series cross-validation: the dataset is divided into multiple sliding time windows (e.g., a window size of 30 days), with each window serving as the validation set sequentially, and the remainder as the training set. The model is trained iteratively (e.g., first training ARIMA using the first 30 days, validating day 31; then sliding the window again, training using days 2-31, validating day 32). Model parameters are optimized using grid search (e.g., adjusting the p, d, and q parameters of ARIMA).

[0046] For example, in predicting factory steam consumption, where the data exhibits strong seasonality, the SARIMA model is selected. During training, 2023 data is used, divided into windows (each window being 30 days): Window 1 (training from January 1st to January 30th, validation on January 31st); Window 2 (training from January 2nd to January 31st, validation on February 1st). SARIMA parameters (such as the seasonality order) are optimized to minimize validation error.

[0047] Model selection is based on data characteristics. Statistical models excel at linear patterns (ARIMA captures trends), deep learning models handle nonlinearity (LSTM captures long-term dependencies), and hybrid models combine advantages (ARIMA-LSTM removes trends before fitting residuals). Cross-validation prevents overfitting because time-series data has autocorrelation, and direct random partitioning could lead to future information leakage. This improves the model's generalization ability.

[0048] S3: Select a suitable single model based on the scene discrimination index, which includes fluctuation amplitude and seasonal cycle; the fluctuation amplitude is calculated by the standard deviation of the data or the range within the moving window, and the seasonal cycle is obtained by detecting the cycle intensity through the autocorrelation function;

[0049] When mapping scenario-based indicators to model fusion selection: for example, if the fluctuation range exceeds a threshold, a deep learning model is preferred; if the seasonal cycle exceeds a threshold, a statistical model is preferred; otherwise, a hybrid model is selected.

[0050] For example, set scenario discrimination indicators (e.g., set the volatility threshold to 0.5, meaning that a volatility standard deviation exceeding 50% is considered high volatility). Dynamically select the rule engine (e.g., decision tree): if the volatility exceeds the threshold, prioritize deep learning models (e.g., LSTM); if there is a clear seasonal cycle (e.g., high autocorrelation function peak), prioritize statistical models (e.g., SARIMA). The rule engine is based on simple logic (e.g., if-else statements).

[0051] In the prediction of factory water consumption, the system calculates the current data fluctuation range (the standard deviation of the past week = 0.6, exceeding the threshold of 0.5), and automatically selects the LSTM model; if a strong daily cycle is detected (the autocorrelation function shows a 24-hour peak), then SARIMA is selected. The rule engine is implemented using a decision tree (e.g., branching rule: if fluctuation > 0.5, use LSTM).

[0052] By dynamically selecting models, the rigidity of single models is addressed. Deep learning models are better able to handle nonlinear abrupt changes (such as equipment failures) when fluctuations are high, while statistical models are more efficient when seasonal cycles are pronounced. The rule engine provides automated decision-making, improving real-time performance.

[0053] S4: Perform model fusion on the prediction results of the selected model;

[0054] During operation: A weighted average method is used for fusion. First, the predicted values ​​of each model are obtained (e.g., the outputs of multiple models selected in S3). Weights are optimized through regression learning: Ridge regression is used, inputting the predicted and true values ​​of each model from historical data, and solving for the weight vector. The optimization formula minimizes prediction error and weight complexity: the goal is to minimize "Σ(true value - weighted predicted value)² + λ × Σweight²". Here, λ (regularization parameter) is selected through cross-validation (e.g., λ = 0.1). Finally, the weighted predicted value is the sum of the predicted values ​​of each model multiplied by their respective weights.

[0055] For example, in power forecasting, S3 selects ARIMA and LSTM models, and inputs their predictions into Ridge regression (trained with historical data) to solve for the weights (e.g., ARIMA weight 0.4, LSTM weight 0.6). "The fusion output = 0.4 × ARIMA prediction + 0.6 × LSTM prediction".

[0056] S5: Evaluate the fusion model and make energy forecasts based on the evaluation results.

[0057] During operation: Evaluation metrics such as Mean Squared Error (MSE) or Mean Absolute Error (MAE) are used. Calculate "MSE: Σ(Predicted value - Actual value)". 2 " / Sample size". If the MSE meets the preset accuracy requirement (e.g., MSE < 0.1), the model is saved for real-time prediction (e.g., deployed to a prediction system). Otherwise, the model or parameters are adjusted (e.g., retrained). For real-time prediction, new data (e.g., daily electricity consumption) is input, and the prediction result (e.g., energy consumption for the next 24 hours) is output.

[0058] After model training, calculate the MSE using test data (e.g., MSE = 0.08 < 0.1) and save the model. For real-time forecasting, for example, input new sensor data and output a forecast of tomorrow's steam consumption to adjust the plant's operating schedule.

[0059] Ensuring model quality through evaluation, such as quantifying errors in MSE or MAE, avoids blind deployment. After saving the model, real-time predictions can be applied to energy management (such as optimized scheduling) to improve business adaptability.

[0060] When implemented by those skilled in the art, standard tools (such as Python's scikit-learn or statsmodels libraries) can be used to automate the process (such as writing scripts to execute S1-S5). For example, in a semiconductor factory energy system, historical energy consumption data can be input and dynamic fusion predictions can be output to improve energy efficiency.

[0061] Further explanation is needed regarding step S1, where energy data (such as electricity consumption data) is cleaned: missing values ​​are filled using interpolation methods, for example, when there are gaps in sensor data, the average value of the preceding and following time points is used; outliers are removed, for example, data points that exceed the normal range are identified and removed using statistical methods. Feature definitions include constructing time-series features (such as calculating the average consumption over the past 7 days as a rolling average), external features (such as adding production plan quantities and ambient temperature), and derived features (such as capturing trends by calculating the magnitude of data changes).

[0062] In step S2, the model type is selected based on data characteristics: if the data exhibits a linear trend (such as energy consumption with obvious seasonality), a statistical model such as ARIMA is selected for training; if the data is non-linear (such as fluctuations caused by equipment start-up and shutdown), a deep learning model such as LSTM is selected; for complex scenarios, a hybrid model such as ARIMA-LSTM is selected. Training employs time-series cross-validation, dividing the data into multiple time windows for sequential training and validation.

[0063] In step S3, the model is dynamically selected: the scenario discrimination index is calculated, the fluctuation amplitude is obtained by the standard deviation of the data within the moving window (e.g., calculating the standard deviation of the data for the most recent 30 days), and the seasonal cycle is detected by the autocorrelation function (e.g., analyzing whether the data has a repeating pattern on a weekly or monthly basis). If the fluctuation amplitude exceeds a threshold (e.g., the standard deviation is greater than a preset value), a deep learning model is selected first; if the seasonal cycle is obvious (e.g., the autocorrelation strength is high), a statistical model is selected first; otherwise, a hybrid model is selected.

[0064] In step S4, the prediction results are fused: a weighted average is used to assign weights to each model.

[0065] In step S5, the fusion model is evaluated, such as by calculating the mean squared error. When the error is below a threshold, the model is stored and used for prediction.

[0066] Adaptive forecasting is achieved through dynamic model selection. For example, in factory energy forecasting, when data experiences sudden fluctuations (such as equipment failure), the system automatically switches to a deep learning model to improve nonlinear fitting capabilities and reduce prediction errors; when data is stable and seasonal (such as monthly consumption), the system prioritizes statistical models to enhance stability. Overall, prediction accuracy is improved, robustness is enhanced, and the risk of single-model failure is avoided. The technical benefits include: improved prediction accuracy, adaptation to data changes, and reduced human intervention.

[0067] Assuming the model is used to predict the power consumption of a semiconductor factory, after data cleaning, feature construction includes adding temperature as an external feature (because high temperatures increase cooling requirements). During training, if the data exhibits significant seasonality (e.g., high consumption in summer), the SARIMA model is chosen; if the data fluctuates greatly (e.g., equipment start-ups and shutdowns), the LSTM model is chosen. Dynamic selection is based on indicators: when fluctuations are large, the weights of the LSTM model increase, resulting in more accurate predictions. After fusion, the evaluation error is small, and the model is used to predict factory load in real time and optimize energy dispatch.

[0068] In this embodiment, in S1, the derived features include time-domain features and frequency-domain features. The time-domain features are calculated by rolling mean or rolling variance, and the frequency-domain features are extracted by Fourier transform or wavelet transform to enhance the model's ability to capture periodicity.

[0069] By using derived features, the model's ability to capture periodic data is enhanced, which solves the prediction bias problem caused by insufficient feature engineering in existing technologies and brings additional accuracy improvement.

[0070] In practice, time-domain features are calculated using either rolling mean or rolling variance: for example, rolling mean calculates the average over a specified period of time (such as 24 hours) to smooth the data and capture trends; rolling variance calculates the magnitude of data variation within that window, indicating volatility.

[0071] Frequency domain features are extracted using Fourier transform or wavelet transform: Fourier transform converts time-domain data into the frequency domain, identifying major periodic components (such as daily or weekly cycles); wavelet transform decomposes the signal into different frequency bands, capturing local periodicity. For example, in energy data, Fourier transform is applied to detect daily consumption peaks as feature input models.

[0072] The aforementioned feature construction techniques enable the model to better identify periodic patterns, such as in seasonal energy consumption, reducing prediction lag errors; frequency domain features help handle sudden noise, improving model robustness. This enhances overall prediction accuracy, especially when data fluctuates.

[0073] For example, when predicting water consumption, the rolling mean is used to calculate the average of the past 7 days as a time-domain feature to show the consumption trend; the Fourier transform extracts the weekly periodic components and inputs them into the LSTM model to make the prediction more accurately match the low consumption pattern on weekends.

[0074] In this embodiment, in S1, the external features include production plan quantity, equipment operating status, and environmental parameters, including temperature and humidity.

[0075] By integrating multi-source external features (such as environmental parameters), the problem of inaccurate predictions caused by neglecting external factors in existing technologies is solved. By enhancing feature diversity, the adaptability of the model to real-world scenarios is improved.

[0076] During implementation, external characteristics include planned production volumes (such as daily factory output), equipment operating status (such as equipment start / stop indicators), and environmental parameters (such as temperature and humidity). These characteristics are acquired through data acquisition systems; for example, temperature sensor data is directly input into the model. During the characteristic definition phase, environmental parameters are used as input variables and correlated with energy data. For example, at higher temperatures, cooling system power consumption increases, and the characteristic values ​​are adjusted accordingly.

[0077] By incorporating external features, the model can more comprehensively reflect business scenarios. For example, when high temperatures cause power peaks, the model can make accurate predictions. Production planning helps correlate output with consumption, reducing errors, improving prediction relevance, and reducing environmental interference.

[0078] For example, in steam consumption prediction, equipment operating status (such as production line running / stopping) is used as a feature. When the equipment stops, consumption decreases, and the model adjusts the prediction accordingly to avoid overestimation.

[0079] In this embodiment, in S2, the statistical model includes ARIMA or SARIMA, the deep learning model includes LSTM, Seq2Seq or Transformer, and the hybrid model includes a combination of ARIMA-LSTM.

[0080] It should be noted that the specific combination and applicable scenarios of model selection can optimize the diversity of the model library, solve the problem of poor generalization ability caused by the single model in the existing technology, and improve the fusion effect.

[0081] In implementation, during step S2, model selection is based on data characteristics: statistical models such as ARIMA are used for linear trend data (e.g., stable seasonal consumption), and SARIMA extended ARIMA handles seasonality; deep learning models such as LSTM are used for nonlinear or long-sequence data (e.g., long-term fluctuations), Seq2Seq handles sequence transformations, and Transformer captures long-distance dependencies; hybrid models such as ARIMA-LSTM combine the advantages of both, for example, ARIMA is first used to handle the linear part, and the residuals are input into LSTM to handle nonlinearity. During training, time series cross-validation ensures generalization. This model combination ensures optimal performance in various scenarios; for example, ARIMA handles simple trends, LSTM handles complex changes, and the fusion reduces overfitting, resulting in lower prediction errors, enhanced model adaptability, and an expanded prediction range.

[0082] For example, when predicting factory power consumption, the linear part of the data is modeled using SARIMA, and the nonlinear fluctuations are modeled using Transformer, resulting in high accuracy after fusion.

[0083] In this embodiment, in S3, the scene discrimination index includes a fluctuation amplitude threshold. When the fluctuation amplitude exceeds the preset threshold, a deep learning model is selected first; when the seasonal cycle characteristics are obvious, a statistical model is selected first.

[0084] It should be noted that by using the direct mapping rule between indicators and model selection, automated decision-making is achieved, solving the inefficiency problem of manual adjustment in existing technologies and improving response speed.

[0085] In step S3, the scenario discrimination indicators include fluctuation amplitude thresholds and seasonal cycle characteristics. Fluctuation amplitude is calculated using the standard deviation of the data (e.g., standard deviation within a moving window). When it exceeds the threshold (e.g., set to 20% of the historical average), a deep learning model is preferred. Seasonal cycle is detected using an autocorrelation function (e.g., calculating the weekly autocorrelation coefficient). When the intensity is high (e.g., coefficient > 0.5), a statistical model is preferred. Rules are automatically triggered based on preset thresholds.

[0086] By using rule-based selection, manual intervention is reduced and efficiency is improved. For example, when data drifts, a quick switch to a suitable model can enhance prediction stability by about 10%. For instance, in steam data, where fluctuations are large (high standard deviation), the system automatically selects LSTM instead of SARIMA, resulting in more accurate predictions.

[0087] In this embodiment, in S3, the selection of the optimal model or model combination is achieved through a rule engine or classifier, wherein the rule engine is built based on decision trees or logistic regression and is used to map scenario indicators to model selection.

[0088] It should be noted that the rule engine is built on decision trees or logistic regression. For example, a decision tree rule engine uses scenario indicators (fluctuation amplitude, seasonal cycle) as input nodes and outputs a model selection; a logistic regression classifier is trained on historical data to predict the best model. In implementation, the indicators are input into the engine, and the decision is automatically output.

[0089] Rule engines make choices more objective and reduce subjective bias; for example, logistic regression learns data patterns, adapts to new scenarios, and improves prediction consistency. For instance, in electricity forecasting, decision tree engines select a hybrid model based on fluctuation and seasonal indicators, making implementation simple.

[0090] In this embodiment, in S4, the weights of the weighted average are obtained through regression learning optimization, including using Ridge regression to fit the relationship between the predicted values ​​of each model and the true values, and solving for the optimal weights; wherein, the optimization formula for Ridge regression is:

[0091] ;

[0092] in, Represents a weight vector, with elements The weights corresponding to the j-th model; Represents the actual value; This represents the weighted average predicted value, calculated using the following formula: ,in Let represent the predicted value of the j-th model for sample i; n represent the number of samples; m represent the number of models; λ represents the regularization parameter used to control the weight complexity; the solution is obtained by cross-validation to determine the optimal λ and . .

[0093] It should be noted that the weights are optimized through regression learning, and regularization is introduced to control the complexity. This solves the coarsening problem of existing simple averaging methods and improves the fusion accuracy and resistance to overfitting.

[0094] In implementation, firstly, the predicted and true values ​​of each model are collected; Ridge regression fits the relationship: minimizing the prediction error and adding a regularization term to control the weight magnitude. The optimization formula minimizes the objective function, including the sum of squared errors and the L2 norm of the weights. During the solution process, cross-validation is used to determine the regularization parameters and weight vector. For example, the weights are iteratively solved on the training set to make the weighted predictions approximate the true values.

[0095] The above methods optimize weight allocation, such as assigning higher weights to high-precision models to reduce fusion errors; regularization prevents overfitting caused by excessive weights, thus improving generalization. This achieves the effects of improved fusion accuracy and enhanced model stability.

[0096] In energy forecasting, LSTM predictions are accurate and have high weights; after solving with Ridge regression, the fusion results are more reliable.

[0097] In this embodiment, in S4, when the model effects are similar, the arithmetic mean or geometric mean is used to perform model fusion on the prediction results of the selected model.

[0098] It should be noted that when models perform similarly (e.g., through similar evaluation errors), an arithmetic mean (sum of predictions from all models divided by the number of models) or a geometric mean (square root of the product of predictions) should be used. For example, when the data is stable and the predictions from all models are consistent, the average can be taken directly. Simple fusion is fast and effective when model performance is balanced, reducing computation time; it is robust, avoids the overhead of weight optimization, and is suitable for resource-constrained scenarios. Technical benefits include ease of implementation and improved efficiency. In short-term water consumption prediction, SARIMA and LSTM errors are similar; using an arithmetic mean for fusion yields stable results.

[0099] In this embodiment, in S5, the evaluation uses mean square error or mean absolute error, and the model is stored when the evaluation result meets the preset accuracy requirements.

[0100] It should be noted that the evaluation uses mean squared error or mean absolute error: the average of the squared differences or the average of the absolute differences between the predicted and actual values ​​is calculated. When the error is below a preset threshold (e.g., mean squared error < 0.1), the model is stored; otherwise, it is retrained. The stored model is then used for real-time prediction.

[0101] Standardized evaluation ensures model quality; for example, deployment is only made when the error is small, avoiding invalid predictions. Storage mechanisms save resources and improve system efficiency. Technical benefits include controllable quality and optimized resources.

[0102] In factory applications, if the mean squared error reaches 0.05, the model is stored; otherwise, the parameters are adjusted and the model is retrained.

[0103] In this embodiment, in S2, the training adopts the time series cross-validation method, which includes dividing the data into multiple time windows, and each window is used as a validation set for iterative training in order to avoid overfitting.

[0104] In implementation, the data is divided into multiple windows (e.g., 5 windows) in chronological order. Each window serves as the validation set, and the rest as the training set. The model is trained iteratively, and the validation set error is evaluated. For example, in energy data, continuous time periods are used to avoid time leakage.

[0105] The methods described above can reduce overfitting, for example, in fluctuating data, ensure model generalization, improve training reliability, and enhance prediction consistency.

[0106] In power prediction training, the data is divided into 5 time windows, and the error of each validation set is small, making the model more stable.

[0107] Example 2

[0108] like Figure 2 As shown, a semiconductor factory energy data prediction system based on the fusion of multiple time series models includes a data processing module, a model training module, a model selection module, a model fusion module, and a model evaluation module.

[0109] The data processing module is used to perform data cleaning and feature definition on energy data. Data cleaning includes removing outliers and filling in missing values, while feature definition includes constructing time-series features, external features, and derived features.

[0110] The model training module is used to select and train multiple single models based on the data characteristics of the energy data processed by the data processing module. The single model can be a statistical model, a deep learning model, or a hybrid model. The training adopts time series cross-validation.

[0111] The model selection module is used to select a suitable single model based on scenario discrimination indicators, which include fluctuation range and seasonal cycle. Fluctuation range is calculated by the standard deviation of the data or the range within the moving window, and seasonal cycle is obtained by detecting the cycle intensity through the autocorrelation function.

[0112] The model fusion module is used to fuse the prediction results of the selected models.

[0113] The model evaluation module is used to evaluate the fusion model and make energy predictions based on the evaluation results.

[0114] The semiconductor factory energy data prediction system based on the fusion of multiple time series models provided in this embodiment has the same implementation principle and technical effect as the method embodiment in Embodiment 1. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in Embodiment 1.

[0115] Example 3

[0116] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a computer to perform the semiconductor factory energy data prediction method based on the fusion of multiple time series models described in Embodiment 1 above.

[0117] Example 4

[0118] An electronic device includes: a memory and a processor, wherein the processor and the memory are connected;

[0119] The memory is used to store programs;

[0120] The processor calls a program stored in the memory to execute the semiconductor factory energy data prediction method based on the fusion of multiple time-series models as described in Embodiment 1.

[0121] It should be noted that the electronic device mentioned may be, but is not limited to, personal computers (PCs), tablet computers, mobile internet devices (MIDs), etc.

[0122] It should be noted that processors, memory, and other components that may be present in electronic devices are electrically connected to each other, directly or indirectly, to enable data transmission or interaction. For example, processors, memory, and other components may be electrically connected to each other via one or more communication buses or signal lines.

[0123] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0124] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0125] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0126] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, laptop, server, mobile phone, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0127] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A semiconductor factory energy data prediction method based on fusion of multiple timing models, characterized by, The method comprises the following steps: S1: data cleaning and feature definition are performed on the energy data, wherein the data cleaning comprises removing outliers and filling missing values, and the feature definition comprises constructing time series features, external features and derived features; the derived features comprise time domain features and frequency domain features, wherein the time domain features are calculated by rolling mean or rolling variance, and the frequency domain features are extracted by Fourier transform or wavelet transform to enhance the periodicity capture of the model; the external features comprise production plan, equipment operation state and environmental parameters, and the environmental parameters comprise temperature and humidity; the outliers are removed by using a Z-score method: calculating the Z-score of the data points, and if the absolute value of the Z-score exceeds 3, the data points are regarded as outliers and removed; the missing values are filled by using a linear interpolation method: for the missing time points, the linear trend of the adjacent data points is used for filling; S2: a plurality of single models are selected and trained based on the data characteristics of the energy data processed in step S1, wherein the single models are statistical models, deep learning models or hybrid models, and time series cross-validation is used for training; the statistical models comprise ARIMA or SARIMA, the deep learning models comprise LSTM, Seq2Seq or Transformer, and the hybrid models comprise an ARIMA-LSTM combination; S3: a suitable single model is selected based on a scene discrimination index, wherein the scene discrimination index comprises fluctuation amplitude and seasonal period; the fluctuation amplitude is calculated by using a standard deviation or a range in a moving window, and the seasonal period is obtained by detecting the periodicity intensity by using an autocorrelation function; the scene discrimination index comprises a fluctuation amplitude threshold, when the fluctuation amplitude exceeds a preset threshold, a deep learning model is preferentially selected; when the seasonal period feature is obvious, a statistical model is preferentially selected; the selected model is realized by using a rule engine or a classifier, wherein the rule engine is constructed based on a decision tree or a logistic regression, and is used for mapping the scene index to the model selection; S4: the prediction results of the selected model are fused; the model fusion is realized by using a weighted average method, and the weight of the weighted average method is obtained by regression learning, comprising fitting the relationship between the predicted values and the true values of each model by using a Ridge regression, and solving the optimal weight; wherein the optimization formula of the Ridge regression is: ； wherein, denotes a weight vector, element corresponds to the weight of the jth model; denotes a true value; denotes a weighted average prediction value, the calculation formula is wherein, denotes the prediction value of the jth model for the ith sample; n denotes the number of samples; m denotes the number of models; λ denotes a regularization parameter, used to control the weight complexity; the solving is to determine the optimal λ and by cross-validation. S5: the fused model is evaluated, and energy prediction is performed based on the evaluation results. 2.A semiconductor factory energy data prediction system based on fusion of multiple time series models, characterized by, The method comprises a data processing module, a model training module, a model selection module, a model fusion module and a model evaluation module. The data processing module is configured to perform data cleaning and feature definition on the energy data, wherein the data cleaning comprises removing outliers and filling in missing values, and the feature definition comprises constructing time-series features, external features, and derived features; the derived features comprise time-domain features and frequency-domain features, wherein the time-domain features are calculated by rolling mean or rolling variance, and the frequency-domain features are extracted by Fourier transform or wavelet transform to enhance the periodicity capture of the model; the external features comprise production plan, equipment operating status, and environmental parameters, and the environmental parameters comprise temperature and humidity; the outliers are removed by using a Z-score method: calculating the Z-score of the data points, and removing the outliers if the absolute value of the Z-score exceeds 3; the missing values are filled in by using a linear interpolation method: for the missing time points, the linear trend of the adjacent data points is used for filling in; The model training module is configured to select and train a plurality of single models according to the data characteristics of the energy data processed by the data processing module, wherein the single models are statistical models, deep learning models, or hybrid models, and the training adopts time series cross-validation; the statistical models comprise ARIMA or SARIMA, the deep learning models comprise LSTM, Seq2Seq, or Transformer, and the hybrid models comprise an ARIMA-LSTM combination; The model selection module is configured to select a suitable single model according to a scene discrimination index, wherein the scene discrimination index comprises fluctuation amplitude and seasonal period; the fluctuation amplitude is calculated by using data standard deviation or range in a moving window, and the seasonal period is obtained by using an autocorrelation function to detect the periodicity; the scene discrimination index comprises a fluctuation amplitude threshold, when the fluctuation amplitude exceeds the preset threshold, the deep learning model is preferentially selected; when the seasonal period feature is obvious, the statistical model is preferentially selected; the optimal model or model combination is selected by using a rule engine or a classifier, wherein the rule engine is constructed based on a decision tree or a logistic regression, and is used for mapping the scene index to the model selection; The model fusion module is configured to perform model fusion on the prediction results of the selected models; the model fusion is performed by using a weighted average method, and the weights of the weighted average method are optimized by regression learning, comprising fitting the relationship between the prediction values and the true values of each model by using Ridge regression, and solving the optimal weights; wherein the optimization formula of the Ridge regression is: ； wherein, denotes a weight vector, elements correspond to the weights of the jth model; denotes the true value; denotes the weighted average prediction value, calculated by wherein, denotes the prediction value of the jth model for sample i; n denotes the number of samples; m denotes the number of models; λ denotes a regularization parameter for controlling the weight complexity; the solution is determined by cross-validation to determine the optimal λ and ; The model evaluation module is configured to evaluate the fused model and perform energy prediction based on the evaluation result.

3. An electronic device, comprising: comprise: a memory and a processor, wherein the processor and the memory are connected; the memory is configured to store a program; the processor is configured to invoke the program stored in the memory to execute the method in claim 1.

4. A computer-readable storage medium, characterized in that, a computer program is stored thereon, and the computer program is configured to execute the method in claim 1 when the computer program is run by a computer.

Images