Semiconductor factory power load forecasting method, system, device and medium

Abstract

The application discloses a semiconductor factory power load prediction method, system, device and medium, and the method comprises the following steps: acquiring historical power load data of a power grid system of a semiconductor factory; determining a lag order according to the historical power load data, and screening input features according to the lag order; adopting a tree model algorithm to fit the input features and power consumption in a future period of time, and constructing the prediction model; and predicting future power load data of the power grid system through the prediction model and outputting a result. According to the application, the feature screening is performed through lag correlation, and the feature is secondarily screened through the feature importance of the tree model, so that the autoregressive prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor manufacturing, in particular to a semiconductor factory power load prediction method, system, device and medium. BACKGROUND

[0002] As the core of the information technology industry, the semiconductor industry has highly precise, automated and complex production processes. The production environment has strict requirements on parameters such as cleanliness, temperature and humidity, and the production equipment (such as photolithography machines, ion implanters, thin film deposition equipment, etc.) has precise structure and high cost. Any minor abnormality or failure can lead to product yield reduction, production interruption and even significant economic losses.

[0003] In the daily operation of a semiconductor factory, power load is one of the key parameters to ensure production stability. Power load not only directly provides energy support for various types of precise production equipment and environmental control systems (such as constant temperature and humidity devices), but also indirectly reflects the equipment operation status and production rhythm - for example, the start of a photolithography machine will cause a short-term increase in local power load. If there is an abnormal fluctuation or insufficient power supply, it may cause equipment downtime, environmental parameter out of control, and thus affect the production process. Therefore, accurate and efficient prediction of the power load of a semiconductor factory is an important prerequisite for early allocation of power resources, identification of potential risks, and protection of production continuity.

[0004] Time series prediction is a common technical means for power load management in semiconductor factories, but due to the complex characteristics of power load data such as high noise, nonlinear trend and multi-period superposition, existing prediction methods still face many challenges in practical application. However, in the context of semiconductor factory power load prediction, current time series prediction techniques have obvious shortcomings: on the one hand, the prediction accuracy is low, making it difficult to accurately capture the complex changes in power load due to factors such as production equipment start-stop, production batch switching, etc., resulting in a large deviation between predicted values and actual power demand, which may cause power resource allocation imbalance; on the other hand, the prediction efficiency is low, and when processing time series data related to power load, it is difficult to effectively eliminate redundant information, increasing the computational complexity and making it difficult to quickly output prediction results, which cannot provide timely support for power dispatching and production adjustment, and is not conducive to the stability of semiconductor factory production. SUMMARY

[0005] To solve the above technical problems, the present application aims to provide a semiconductor factory power load prediction method that determines the input parameters adaptively through lag correlation and performs nonlinear fitting using ensemble learning.

[0006] The semiconductor factory power load prediction method of the present application comprises:

[0007] Obtaining historical power load data of a power grid system of a semiconductor factory;

[0008] Determining a lag order according to the historical power load data, and screening input features according to the lag order;

[0009] Fitting the input features and power consumption in a future period of time by using a tree model algorithm to construct the prediction model;

[0010] Predicting future power load data of the power grid system by using the prediction model and outputting a result.

[0011] The step of screening input features according to the lag order comprises:

[0012] Calculating a correlation between a sequence and a sequence lagging the sequence by n orders, wherein n is a length of the sequence x;

[0013] Screening the sequence lagging the sequence by n orders according to a predetermined threshold to obtain the input features.

[0014] In the step of calculating the correlation between the sequence and the sequence lagging the sequence by n orders, the following is met:

[0015]

[0016] wherein n is the length of the sequence x, d is the lag order, is a correlation coefficient of the sequence x and the sequence x d , is an original sequence, is a sequence lagging d orders, is an i-th value in the sequence x, is an i+d-th value in the sequence x, is an i-th value in the sequence , is a mean value of the sequence x, is a mean value of the sequence .

[0017] In the step of obtaining the input features, the lag order with a correlation ≥ 0.7 is selected as an input feature for training the tree model.

[0018] The step of training the tree model algorithm on the input features and the dependent variable comprises:

[0019] Fitting the input features and the dependent variable by using the tree model algorithm;

[0020] Screening the input features according to importance, removing the input features with an importance of 0 and refitting until all the input feature variables have a weight of 0.

[0021] In the step of fitting the input features and the dependent variable by using a tree model algorithm, the input features and the dependent variable are fitted by a LightGBM method or an Xgboost method.

[0022] In the step of screening the input features according to importance, removing the input features with an importance of 0 and re-fitting until the weight of all the input feature variables is not 0, the importance of the input features is calculated by a SHAP method.

[0023] The semiconductor factory power load prediction system comprises:

[0024] An acquisition module acquires historical power load data of a power grid system of a semiconductor factory.

[0025] A calculation module determines a lag order according to the historical power load data, and screens input features according to the lag order.

[0026] A construction module fits the input features and power consumption in a future period of time by using a tree model algorithm, and constructs the prediction model.

[0027] A prediction module predicts future power load data of the power grid system by using the prediction model and outputs a result.

[0028] The semiconductor factory power load prediction device comprises a memory storing a semiconductor factory power load prediction method program and a processor for running the semiconductor factory power load prediction method program, and the semiconductor factory power load prediction method program is configured to realize the steps of the semiconductor factory power load prediction method.

[0029] The computer readable storage medium stores the semiconductor factory power load prediction method program, and the semiconductor factory power load prediction method program realizes the steps of the semiconductor factory power load prediction method when executed by the processor.

[0030] The calculation module realizes lag correlation feature screening, retains high-correlation historical data, changes one-dimensional prediction (today's power consumption prediction tomorrow's power consumption) into multi-dimensional prediction (33 days of power consumption in the past 400 days prediction tomorrow's power consumption), and reduces the calculation complexity. The construction module performs secondary screening on the features by using the feature importance of the tree model, removes some redundant features, improves the autoregressive prediction accuracy, and further improves the prediction efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0031] Other features, objects, and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments thereof, when read in connection with the following accompanying drawings.

[0032] Figure 1 A flow chart of a semiconductor factory power load prediction method in one embodiment;

[0033] Figure 2 A data sample chart in one embodiment;

[0034] Figure 3 A lag correlation fold line chart in one embodiment;

[0035] Figure 4 A correlation bar chart of remaining lag data and target in one embodiment;

[0036] Figure 5 A prediction data and true data comparison chart in one embodiment. DETAILED DESCRIPTION

[0037] The semiconductor factory power load prediction method of the present application is described in further detail below in conjunction with the accompanying drawings. In the following detailed description of the application, certain exemplary embodiments of the application are described in terms of the procedures, systems, and methods detailed herein. Description of the application in a particular aspect with reference to a specific embodiment is intended to include all combinations and permutations of one or more features specific to that embodiment with one or more features from other like embodiments. No limitation on the scope of the application is intended by reference to a specific aspect, embodiment, or method.

[0038] Time series prediction is a key technology in the fields of finance, meteorology, energy, etc. For the above scenarios, the semiconductor factory power load prediction method of the present application adaptively determines the features of the input through lag correlation, then performs nonlinear fitting using ensemble learning, and further filters the features according to the feature importance.

[0039] The steps of the present application include:

[0040] ①Adaptively determine the features of the input through lag correlation

[0041] Adaptively determine the features of the input through lag correlation,

[0042] The correlation of lag n order, expressed as:

[0043]

[0044] Where n is the length of the sequence x, d is the lag order, is the correlation coefficient of the sequence x and the sequence xd, is the original sequence, is the sequence of lag d order, the i-th value in the sequence x, the i+d-th value in the sequence x, the i-th value in the sequence the i-th value in the sequence x, the mean of the sequence x, the mean of the sequence the mean of the sequence x.

[0045] Determine n according to the frequency of the data. If the data is daily, then n = 400; if the data is weekly, then n = 60.

[0046] Keep the order of correlation greater than or equal to 0.7 as the input feature of the autoregressive model. In the embodiment of the application, the lag order kept is 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380. Then when predicting the value at time t, the input features are the values at times t-1, t-2, t-3, t-4, t-5, t-6, t-7, t-8, t-9, t-10, t-11, t-12, t-13, t-14, t-362, t-363, t-364, t-365, t-366, t-367, t-368, t-369, t-370, t-371, t-372, t-373, t-374, t-375, t-376, t-377, t-378, t-379, t-380.

[0047] ②Nonlinear fitting is performed using a tree model, and the features are again screened according to feature importance.

[0048] Nonlinear fitting is performed using ensemble learning, wherein fitting is learning the input features and the target to be predicted, and the error is minimized during the learning process. In the present application, fitting optimizes the model parameters by minimizing the loss function (such as mean square error MSE, cross entropy) to minimize the deviation between the predicted value and the true value. The tree model selects the optimal feature and threshold for feature division by recursion, and each node represents a division decision, and the leaf node stores the category or numerical result. For the tree model, the training is similar to that of a neural network. The parameters are again screened according to feature importance:

[0049] All independent variables and dependent variables are fitted using ensemble learning such as LightGBM and Xgboost, the independent variables with a feature importance of 0 are removed, and refitting is performed until the weights of all independent variables are not 0.

[0050] The feature importance is calculated according to a SHAP (SHapley Additive exPlanations) method.

[0051] Embodiment

[0052] Step 1: Obtain historical data of a system to be predicted to build a model training data set.

[0053] In this embodiment, the data used for rationality verification is the power load of a certain province from August 1, 2019 to December 31, 2021. The data is daily frequency data, one piece per day. An example of the data is shown in Figure 2 .

[0054] Step 2: Determine the parameters of the input through lag correlation self-adaptation.

[0055] This step 2 includes three sub-steps.

[0056] Step 2.1: Determine the order of the lag according to the frequency of the data. The lag of the daily frequency data is 400 orders, which can include one year of data; the lag of the weekly frequency data is 60 orders, since one year is more than 52 weeks, taking 60 weeks contains one year of data.

[0057] Step 2.2: Calculate the correlation of the lag n orders, the correlation satisfies:

[0058]

[0059] Wherein, n is the length of the sequence x, d is the lag order, is the correlation coefficient of the sequence x and the sequence xd, is the original sequence, is the sequence of lag d orders, is the i-th value in the sequence x, is the i+d-th value in the sequence x, is the i-th value in the sequence , is the mean of the sequence x, is the mean of the sequence .

[0060] Step 2.3: Keep the order whose correlation is greater than or equal to 0.7 (it is generally considered that the correlation is greater than 0.7 is high correlation) as the input variable of the autoregressive model.

[0061] In this embodiment, the daily frequency data is used, so the lag order is 400 orders, and the lag correlation is as shown in Figure 3 . According to the correlation threshold, the lag orders of the parameters retained are 1-14 and 362-380, a total of 33, and the correlation column chart is as shown in Figure 4The results are shown in the following table.

[0062] Step 3, build a tree model (this embodiment uses a boosting tree), and further screen the features according to feature importance.

[0063] In this embodiment, the tree model is used to fit the features and the target, and the data from August 1, 2019 to June 30, 2021 is used as training data, and the data from July 1, 2021 to December 31, 2021 is used as test data to compare the performance of multiple models.

[0064] In this embodiment, the importance of each feature is shown in the following table:

[0065] Table 1: Importance of Specialties

[0066]

[0067] The importance of the features is obtained by SHAP value. SHAP value (SHapley Additive exPlanations) is a tool for explaining the prediction results of machine learning models, which can quantify the contribution of individual features to model prediction. The core idea is to decompose the model prediction results into the cumulative value of each feature influence, so as to intuitively show the importance of the features. SHAP value calculates the marginal contribution of each feature in model prediction, and decomposes the prediction results into the cumulative value of each feature influence. For example, in tree models (such as XGBoost), SHAP value calculates the importance of features in tree node partitioning; in permutation importance method, the model performance change is observed by randomly shuffling feature values.

[0068] The specific way is: 1. Calculate the difference between the prediction output containing the feature and the prediction output not containing the feature for each feature. 2. Weighted average of the prediction difference of all possible feature subsets, so as to get the feature importance of each feature.

[0069] Through feature importance, the features lagging 370 days and lagging 379 days are removed, leaving 31 features.

[0070] Step 4, predict future trends.

[0071] In this embodiment, the electricity consumption from July 1, 2021 to December 31, 2021 is predicted.

[0072] It should be noted that in order to avoid using future data when predicting, if the independent variable at a certain time point needs to use data after July 2021, the previously predicted data will be used instead.

[0073] Prediction results Figure 5The figure shows that the model proposed by the present application has the best prediction ability, the Lasso algorithm is the second, and the ARIMA algorithm is the worst. The R square also proves this point. The R square of the model proposed by the present application is 0.624, the R square of the Lasso algorithm is 0.168, and the R square of the ARIMA algorithm is-6.272.

[0074] As can be seen from the figure, the model proposed by the present application has the best prediction ability, the Lasso algorithm is the second, and the ARIMA algorithm is the worst. The R square also proves this point. The R square of the model proposed by the present application is 0.624, the R square of the Lasso algorithm is 0.168, and the R square of the ARIMA algorithm is-6.272.

[0075] The above has specifically described the preferred embodiments of the present application, but the present application is not limited to the embodiments, and those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the present application, and these equivalent modifications or replacements are all included in the scope of the present application.

Claims

1. A method for predicting an electrical load of a semiconductor factory, characterized by, The method comprises: obtaining historical power load data of a power grid system of a semiconductor factory; determining a lag order according to the historical power load data, and screening input features according to the lag order; fitting the input features and power consumption in a future period of time by using a tree model algorithm to construct a prediction model; predicting future power load data of the power grid system by using the prediction model and outputting a result; the step of screening input features according to the lag order comprises: calculating the correlation between a sequence and a sequence lagged by n orders of the sequence, wherein n is the length of the sequence x; screening the sequence lagged by n orders of the sequence according to a predetermined threshold to obtain the input features; in the step of calculating the correlation between a sequence and a sequence lagged by n orders of the sequence, the following condition is met: where n is the length of the sequence x, d is the lag order, is the correlation coefficient of the sequence x and the sequence x d , is the original sequence, is the sequence with a lag of d, is the i-th value in the sequence x, is the i+d-th value in the sequence x, is the i-th value in the sequence , is the mean of the sequence x, is the mean of the sequence ; in the step of obtaining input features, the lag order with a correlation greater than or equal to 0.7 is selected as an input feature for training the tree model; the step of training the input features and the dependent variable by using the tree model algorithm to construct the tree model comprises: fitting the input features and the dependent variable by using the tree model algorithm; screening the input features according to importance, removing the input features with an importance of 0 and refitting until the weight of all the input feature variables is not 0; in the step of fitting the input features and the dependent variable by using the tree model algorithm, the input features and the dependent variable are fitted by using a LightGBM method or an Xgboost method; in the step of screening the input features according to importance, removing the input features with an importance of 0 and refitting until the weight of all the input feature variables is not 0, the importance of the input features is calculated by using a SHAP method.

2. A semiconductor factory power load forecasting system comprising the semiconductor factory power load forecasting method according to claim 1, characterized by, The method comprises: an obtaining module configured to obtain historical power load data of a power grid system of a semiconductor factory; a calculating module configured to determine a lag order according to the historical power load data, and screen input features according to the lag order; a constructing module configured to fit the input features and power consumption in a future period of time by using a tree model algorithm to construct a prediction model; a predicting module configured to predict future power load data of the power grid system by using the prediction model and output a result.

3. A semiconductor factory power load forecasting device, characterized by comprising: The method comprises: a memory storing a semiconductor factory power load prediction method program and a processor configured to execute the semiconductor factory power load prediction method program, wherein the semiconductor factory power load prediction method program is configured to implement the steps of the semiconductor factory power load prediction method according to claim 1.

4. A computer readable storage medium, characterized in that, A semiconductor factory power load prediction method program is stored on a computer readable storage medium, and the semiconductor factory power load prediction method program is executed by a processor to implement the steps of the semiconductor factory power load prediction method according to claim 1.

Images