Material demand analysis and prediction method considering uncertainty under data driving

Abstract

The invention discloses a data-driven material demand analysis and prediction method considering uncertainty, and the method comprises the following steps: (1) dividing business expansion materials into continuous demand materials and intermittent demand materials based on historical demand data; (2) respectively carrying out k-means clustering analysis on the continuous demand materials and the intermittent demand materials; (3) according to the clustering result, a time sequence model and a deep learning model are selected for each subclass of materials for demand prediction, the time sequence model is an autoregressive integral moving average model ARIMA, and the deep learning model is a long short-term memory network LSTM; comparing prediction effects of different models through evaluation indexes, selecting an optimal prediction result as a final demand prediction value, and applying the result to a material purchasing decision; according to the invention, a scientific and reasonable decision basis is provided for business expansion project material purchase management.

Description

Technical Field

[0001] The present invention relates to the technical field of material demand, and in particular to a data-driven material demand analysis and forecasting method taking uncertainty into consideration. Background Art

[0002] Traditional statistical methods, machine learning, and deep learning algorithms can be trained and learned through historical data on power grid material demand. However, there are also some shortcomings in directly using these methods when predicting power material demand. First, the factors that affect the material demand of power grid projects are extremely complex and constantly changing. For different power material demands, their historical usage also varies greatly from each other. In addition, for industry expansion materials, there are fluctuations in demand and high response requirements. Therefore, choosing a fixed method to directly predict is not effective, and it is not possible to make targeted predictions based on the characteristics of the materials, which leads to a decrease in the accuracy of the prediction and makes it difficult to meet the standards for actual production stocking. Summary of the invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a data-driven material demand analysis and forecasting method taking uncertainty into account, so as to solve the problem of forecasting the demand for business expansion project materials that are greatly affected by external factors such as seasonal changes and economic development under uncertainty. After classifying the material types, combined with the distribution characteristics of material demand, the most appropriate ARIMA time series model and LSTM model are selected for forecasting analysis, and the LSTM model considers not only the time series but also external factors, thereby reducing the forecast error of each type of material demand and improving the forecast efficiency and accuracy.

[0004] Technical solution: The data-driven material demand analysis and forecasting method considering uncertainty described in the present invention comprises the following steps: (1) Based on historical demand data, business expansion materials are divided into continuous demand materials and intermittent demand materials; among them, continuous demand materials are materials with large outbound quantities or small outbound quantities but scattered peak distribution, and intermittent demand materials are materials with small outbound quantities and concentrated peak distribution; (2) performing k-means cluster analysis on the continuously demanded materials and intermittently demanded materials respectively; (3) According to the clustering results, for each sub-category of materials, the time series model or deep learning model is selected for demand forecasting. The time series model is the autoregressive integrated moving average model (ARIMA), and the deep learning model is the long short-term memory network (LSTM). (4) Compare the prediction effects of different models through evaluation indicators, select the optimal prediction result as the final demand prediction value, and use the result for material procurement decisions.

[0005] Furthermore, in step (1), the business expansion materials include transformers, distribution boxes, smart meters, overvoltage protection equipment, cable accessories, and distribution network automation terminals, and the demand classification further includes the following rules: if there are continuous non-zero demand records for the materials in the historical data that exceed a preset threshold, then it is determined to be a continuous demand material; if the demand record of the materials presents a periodic zero value and the peaks are concentrated, then it is determined to be an intermittent demand material; wherein the threshold value range is 3-6 months, and the preset period value range is 15-30 days.

[0006] Furthermore, in step (3), the parameter selection of the ARIMA model is determined by optimizing the AIC index; the structure of the LSTM model includes a forget gate, an input gate, and an output gate, the dimension of its input layer is consistent with the length of the historical demand sequence, the number of hidden layer neurons is 64, and the output layer is a fully connected layer.

[0007] Furthermore, the training process of the LSTM model includes: normalizing the historical demand data and generating sequence data according to the time window; using the Adam optimizer to update the parameters, and the loss function is the mean square error; setting the Early Stopping mechanism to prevent overfitting.

[0008] Furthermore, the evaluation indicators in step (4) include mean square error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE), and the optimal prediction model is selected by comparing the indicator values ​​of different models.

[0009] Furthermore, the input data for cluster analysis in step (2) include: time series statistical characteristics of material demand, including mean, variance, and maximum continuous non-zero period; time domain characteristics of demand distribution, including peak interval, peak duration, and zero value ratio; the number of clusters is determined by the silhouette coefficient method, the number of clusters for continuously demanded materials is 4, and the number of clusters for intermittently demanded materials is 2.

[0010] Furthermore, in step (3), the LSTM model introduces external factor data, including GDP and electricity consumption data, to improve the prediction accuracy.

[0011] Furthermore, in step (4), the final demand forecast value is used to generate a material procurement plan, and the procurement strategy is dynamically adjusted according to the forecast error.

[0012] Beneficial effects: Compared with the existing technology, this model adopts a clustering-ARIMA-LST combined forecasting model. Starting from the type of materials, it is classified according to the data characteristics of historical demand. Combined with the time series influencing factors of each type of materials, the ARIMA model is selected or GDP and electricity consumption data are added to carry out forecasting analysis of the LSTM model. This can improve the accuracy of the forecast, realize quantitative and accurate analysis and forecast of fluctuations in demand for business expansion materials, and provide a scientific and reasonable decision-making basis for the procurement management of materials for business expansion projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0014] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0015] like Figure 1 As shown, an embodiment of the present invention provides a data-driven material demand analysis and forecasting method considering uncertainty, comprising the following steps: (1) Material classification and cluster analysis Step 1.1: Demand classification rules: If there are continuous non-zero demand records exceeding the preset threshold in the historical data of the material, it is determined to be a continuously demanded material; if the demand record of the material presents periodic zero values ​​and the peaks are concentrated, it is determined to be an intermittently demanded material.

[0016] Step 1.2: k-means cluster analysis: Input data include: time series statistical characteristics of material demand, including mean, variance, and maximum continuous non-zero period; time domain characteristics of demand distribution, including peak interval, peak duration, and zero value ratio; the silhouette coefficient method is used to measure the closeness of each sample to its cluster and the degree of separation from other clusters, with a value range of , the closer the value is to 1, the better the clustering effect is. By calculating the average silhouette coefficient under different numbers of clusters, the number of clusters corresponding to the maximum average silhouette coefficient is selected as the value of n_clusters. The random_state parameter is used to set the seed of the random number generator to ensure the reproducibility of the clustering results. Its value can be any integer, and a fixed integer is usually selected (generally 42). Therefore, the parameter selection for continuous demand materials is: `n_clusters=4`, `random_state=42`. The parameter selection for intermittent demand materials is: `n_clusters=3`, `random_state=42`. Output result: continuous demand materials are clustered into 4 categories; intermittent demand materials are clustered into 2 categories.

[0017] (2) Construction of combined prediction model Step 2.1: ARIMA model time series model For the demand data of business expansion materials, there may be a lack of characteristic variables that can characterize the demand response value, so a time series forecasting model is needed. The time series model is a type of statistical model specifically used to analyze time series data. Time series data is a series of data points arranged in chronological order. The model only needs to input the target response value. These models can capture the temporal trend, seasonal pattern, periodicity and other characteristics of the data, and are used to predict future values. This model chooses to use the autoregressive integrated moving average model (ARIMA, Autoregressive Integrated Moving Average) to predict the demand for business expansion materials. This method can effectively deal with non-stationary sequences. By using the difference operation of appropriate order to achieve post-differentiation stability for the non-stationary sequence, an ARIMA model can be constructed for the post-differentiation sequence.

[0018] ARIMA is a commonly used time series analysis and forecasting method. It combines the characteristics of the autoregressive (AR) model, the difference (I) and the moving average (MA) model to capture the autocorrelation, trend and seasonality of the data. The model with the following structure is called the summed autoregressive moving average model, abbreviated as Model: ; The establishment of ARIMA model usually includes three main steps: determining the model order, estimating model parameters and model diagnosis. First, determine the order of the model. ARIMA model consists of three parameters: p, d and q, which represent the autoregressive order (AR), the difference order (I) and the moving average order (MA) respectively. By observing the autocorrelation diagram (ACF) and partial autocorrelation diagram (PACF) of time series data, the order of the model can be preliminarily determined. The second step is to estimate the model parameters. Using methods such as Maximum Likelihood Estimation, the parameters of AR, I and MA are estimated based on historical data. The last step is model diagnosis. By testing the residual sequence, it is determined whether the model fitting effect and the residual sequence meet the model assumptions. Commonly used diagnostic methods include checking the autocorrelation of residuals, white noise test, normality of residuals, etc. ARIMA model can be applied to various types of time series data, including economic data, meteorological data, stock prices, etc. It can be used to predict future values, analyze the trend and seasonality of data, and smooth data. The combination with the smallest AIC value is selected by using the AIC indicator. , which is selected as the model for fitting, so the ARIMA model used for business expansion material demand forecasting is .

[0019] Step 2.2: LSTM Model: In the data-driven demand forecasting method, deep learning can show better results in the processing and forecasting scenarios with large sample data. With its ability of adaptive feature extraction, it shows more outstanding performance in learning, monitoring and forecasting, making the processing process more intelligent. There are many deep learning algorithm models, among which LSTM is an improved recurrent neural network. It adds a long-term memory storage structure on the basis of the recurrent neural network, overcomes the memory defects of the recurrent neural network, and can effectively process time series data.

[0020] Recurrent neural networks (RNNs) perform well in processing sequence data. In traditional neural network models, from the input layer to the hidden layer and then to the output layer, the layers are fully connected, and the nodes between each layer are disconnected. RNN is derived from ordinary neural networks. Specifically, the hidden layer neurons of RNN are connected. When RNN processes time series data, the hidden layer neurons are not only connected to the neurons in the previous layer, but also to other hidden layer neurons. In other words, there is also information transmission between the hidden layer neurons, which makes RNN more applicable to time series data with dependencies between the previous and the next. It is precisely because the structure of hidden layer neurons transmitting information is added to RNN that it has a better effect on time series data processing.

[0021] Similar to RNN, the hidden layers of LSTM network are also interconnected, and the hidden layers can transmit information in a specific direction. The difference is that in LSTM, the unit state that can memorize information for a long time is introduced. Indicates that the unit output is To indicate that Mainly used for long-term memory of information, Mainly used for short-term memory of information. In addition, LSTM contains three gate structures: forget, input, and output.

[0022] The first is the forget gate, which is used to pass the input information. To calculate the coefficient of the amount of information retained. The calculation formula is: ; in and represents the weight and bias parameters of the forget gate, for The function is to scale the calculated value to the interval [0,1].

[0023] The second is the input gate, which is used to determine the new cell state ,in Function Output Decide what information to update, Function determines the update initial information , the update formula is as follows: ; ; ; in, , , , Represents weight and bias parameters.

[0024] The last one is the output gate, which gets the output based on the cell state. Function Output Determines the amount of information output, and The unit states acted upon by the function are multiplied to form the final output of this unit. The calculation formula is as follows: ; ; in, and Represents the weight and bias parameters of the output gate.

[0025] Analyze the demand for each type of material. First, draw a line graph of the data over time to observe whether there is an obvious trend (increase / decrease) and seasonality (cyclical fluctuation). Then use the unit root test to test whether the data is stable (that is, the mean and variance do not change over time). If the p value is <0.05, reject the null hypothesis (the data is not stable, that is, there is a time trend). Finally, divide the training set and the test set, and try to fit the data with the ARIMA model. If the model can effectively perform predictive analysis, the data is related to time. Otherwise, this type of material is not related to the time series. Combined with the characteristic analysis of materials on time series, the ARIMA time series model is selected for materials that are greatly affected by time changes, otherwise the LSTM model is selected. External factors such as GDP and electricity consumption are considered in the selection of the LSTM model to improve the accuracy of demand forecasting.

[0026] (3) Optimization and output of prediction results.

[0027] The mean absolute percentage error (MAPE) is used to evaluate the prediction effect. The reason why the mean absolute percentage error can describe the accuracy is that the mean absolute percentage error itself is often used as a statistical indicator to measure the accuracy of predictions, such as the prediction combination of time series. The formula is as follows: ; The model is trained by the mean absolute percentage error to obtain the optimal parameters for prediction analysis and to evaluate the prediction results. , then consider changing the prediction method, compare the prediction errors of ARIMA and LSTM methods, and select the best prediction result.

Claims

1. A data-driven material demand analysis and forecasting method considering uncertainty, characterized in that: The following steps are involved: (1) Based on historical demand data, business expansion materials are divided into continuous demand materials and intermittent demand materials; among them, continuous demand materials are materials with large outbound quantities or small outbound quantities but scattered peak distribution, and intermittent demand materials are materials with small outbound quantities and concentrated peak distribution; (2) performing k-means cluster analysis on the continuously demanded materials and intermittently demanded materials respectively; (3) According to the clustering results, a time series model and a deep learning model are selected for each subcategory of materials to predict demand. The time series model is the autoregressive integrated moving average model (ARIMA), and the deep learning model is the long short-term memory network (LSTM). (4) Compare the prediction effects of different models through evaluation indicators, select the optimal prediction result as the final demand prediction value, and use the result for material procurement decisions.

2. According to claim 1, a data-driven material demand analysis and forecasting method considering uncertainty is characterized in that: In step (1), the business expansion materials include transformers, distribution boxes, smart meters, overvoltage protection equipment, cable auxiliary materials, and distribution network automation terminals.

3. According to claim 1, a data-driven material demand analysis and forecasting method considering uncertainty is characterized in that: In step (1), the demand classification further includes the following rules: if there are continuous non-zero demand records exceeding a preset threshold in the historical data of the material, it is determined to be a continuously demanded material; if the demand record of the material presents a periodic zero value and the peaks are concentrated, it is determined to be an intermittently demanded material; wherein the threshold value ranges from 3 to 6 months, and the preset period ranges from 15 to 30 days.

4. According to claim 1, a data-driven material demand analysis and forecasting method considering uncertainty is characterized in that: In step (2), the input data of cluster analysis include: the time series statistical characteristics of material demand include mean, variance, and maximum continuous non-zero period; the time domain characteristics of demand distribution include peak interval, peak duration, and zero value ratio; the number of clusters is determined by the silhouette coefficient method, and the number of clusters for continuously demanded materials is 4; the number of clusters for intermittently demanded materials is 2.

5. According to claim 1, a data-driven material demand analysis and forecasting method considering uncertainty is characterized in that: In step (3), the parameter selection of the ARIMA model is determined by optimizing the AIC index; the structure of the LSTM model includes a forget gate, an input gate, and an output gate. The dimension of its input layer is consistent with the length of the historical demand sequence, the number of neurons in the hidden layer is 64, and the output layer is a fully connected layer.

6. The data-driven material demand analysis and forecasting method considering uncertainty according to claim 5, characterized in that: The training process of the LSTM model includes: normalizing the historical demand data and generating sequence data according to the time window; using the Adam optimizer to update the parameters, and the loss function is the mean square error; setting the EarlyStopping mechanism.

7. The data-driven material demand analysis and forecasting method considering uncertainty according to claim 5, characterized in that: In step (3), the LSTM model introduces external factor data including GDP and electricity consumption data.

8. The data-driven material demand analysis and forecasting method considering uncertainty according to claim 1, characterized in that: The final demand forecast value in step (4) is used to generate a material procurement plan and dynamically adjust the procurement strategy based on the forecast error.

9. The data-driven material demand analysis and forecasting method considering uncertainty according to claim 1, characterized in that: The evaluation indicators in step (4) include mean square error (MSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), and the optimal prediction model is selected by comparing the indicator values ​​of different models.

Images