Otomotiv döküm parça talep tahmininin yapay sinir ağları ile modellenmesi ve arıma yöntemi ile karşılaştırmalı analizi = Modeling automotive casting part demand forecasting with artificial neural networks and comparative analysis with arima method

Abstract

Şirketler kar edebilmek, sürekliliğini korumak ve büyümek için bazı politikalar uygularlar. Bu politikalara stok yönetimi, üretim planlaması, tedarik zinciri yönetimi ve maliyet yönetimi örnek verilebilir. Uygulanan bu politikaların temelini talep tahmini oluşturur. Otomotiv sektörü geçen her günde teknolojik gelişmeler ile daha da gelişmektedir. Otomotiv sektörünün gelişmesi uygulanan talep tahmini yöntemlerinin de gelişmesi gerektirdiğini ortaya çıkarmıştır. Döküm sektöründe faaliyet gösterip otomotiv sektörünün en önemli tedarikçilerinden birisi olan firmada müşteri ihtiyaçlarını karşılamak ve üretimi yönlendirmek için teknoloji ve üretimi birleştiren şirket için yeni bir talep tahmini yöntemi olan bu çalışma yapılmıştır. Döküm sektörü çok fazla girdiye, üretim olarak dar boğaza, yüksek üretim çeşitliliğine sahiptir. Hem üretim çeşitliliğini hem de kapasite kısıtlarını aynı anda yönetmek ayrıca müşteri beklentilerini tam zamanında karşılamak oldukça zordur. Bu zorluğun önüne geçmekte talep tahmini önemli bir araçtır. Literatürde cam sektörü, ekonomik kurlar, enerji harcamaları, giyim sektörü, beyaz eşya sektörü, yiyecek içecek tüketimleri, otomotiv satışları, demir çelik sektörleri gibi farklı çalışma alanlarında talep tahmini çalışmaları yapılmıştır. Yapılan talep tahmini çalışmalarında yapay sinir ağı, makine öğrenmesi, zaman serisi yöntemleri, istatiksel yöntemler gibi farklı talep tahmini yöntemleri kullanılmıştır. Bu çalışmada yapay sinir ağının farklı katman sayıları, farklı nöron sayıları ve farklı öğrenme algoritmaları kullanılarak talep tahmini yapılmış, ARIMA yöntemiyle karşılaştırılarak tahmin performansı en iyi talep tahmini yöntemi ve modeli belirlenmiştir. Yapay sinir ağı (YSA) tahminlerinin tahmin performansları ortalama mutlak hata (MAE), R Kare, kök ortalama kare hata (RMSE), mutlak ortalama yüzde hata (MAPE) analizleri ile karşılaştırılmıştır. En iyi yapay sinir ağına R Kare analizine göre karar verilmiştir. Yapay sinir ağı yönteminin sonuçları ARIMA yönteminin sonuçlarıyla karşılaştırılmıştır. ARIMA modellerinin tahmin performansları Akaike Bilgi Kriteri kullanılarak karşılaştırılmıştır. Seçilen parçanın satışını etkileyen kriterler hem uzman görüşleri alınarak hem de literatür taraması yapılarak kamyon üretim adeti, elektrik birim fiyatı, Euro kuru, yurtiçi üretici fiyat endeksi, genel tüketici fiyat endeksi, brent petrol varil fiyatı, hurda fiyatları, pik fiyatları ve gayri safi yurt içi hasıla olarak belirlenmiştir. Çıktı olarak ise parça aile grubuna ait satış adetleri kullanılmıştır. Çalışmada kullanılan veri ise girdi ve çıktı kriterlerine ait 2017 ve 2022 yılları arasındaki aylık verileridir. Yapay sinir ağı yöntemiyle yapılan talep tahmini çalışmalarından R Kare hata analiz yöntemine göre en iyi tahmin performansına sahip çalışma traincgb öğrenme algoritmalı, 2 katman 10 nöron, tansig aktivasyon fonksiyonlu yapay sinir ağı modeli olduğu sonucuna ulaşılmıştır. Ayrıca yapay sinir ağı yöntemiyle zaman serisi analizi yöntem sonucunu karşılaştırmak için satış verilerine ARIMA yöntemi uygulanmıştır. ARIMA modelleri arasından Akaike Bilgi Kriteri değeri baz alınarak en iyi ARIMA modeli ARIMA (1,1,1) seçilmiştir. Yapay sinir ağı ile ARIMA (1,1,1) modeli R Kare hata analiz yöntemiyle karşılaştırılarak yapay sinir ağı modelinin tahmin performansının ARIMA modelinden daha yüksek olduğu sonucuna ulaşılmıştır.

Companies implement specific policies to make profits, maintain continuity, and grow. These policies include inventory management, production planning, supply chain management, and cost management. Demand forecasting forms the basis of these policies. The automotive industry is developing more and more with technological developments every day. The development of the automotive industry has revealed that the applied demand forecasting methods also require development. This study, a new demand forecasting method, was conducted to meet customer needs and direct production in a company that operates in the casting industry and is one of the most important automotive industry suppliers. For the study, the critical parts, which have uncertain customer demand, create a bottleneck in production, and fill the capacities, were determined by taking expert opinions. The production stages of the part can be summarized as follows: First, a mold suitable for the desired shape is prepared. At the same time, while the metal suitable for the intended use of the part is prepared in the melting furnaces, special cores are produced to create cavities inside the part. After preparing the mold, metal, and core, the sand mold casting method creates the part. However, at this stage, the production of the part has been completed, but its operations still need to be completed. The part is cleaned of sand and made to meet customer requirements. The part can then be shipped to the customer. While the casting industry is environmentally friendly by ensuring the reuse of recyclable materials, it is advantageous in reducing costs in mass production and the sensitivity of the parts. The casting industry is a sector that feeds the automotive industry in a wide range. The casting method produces engine blocks, cylinder heads, transmission, suspension, body, brake, and clutch parts used in the automotive industry. The casting industry has much input. Additionally, the variety of parts is very high. This causes bottlenecks in production. It is complicated to manage production diversity and capacity constraints simultaneously and meet customer expectations on time. Demand forecasting is an essential tool in avoiding this difficulty. A high forecast performance of the demand forecast result enables the company to get ahead of its competitors, develop, grow, and become trusted by customers. When the literature is examined, it has been observed that demand forecasting studies have been carried out in different fields of study, such as the glass sector, economic exchange rates, energy expenditures, clothing sector, white goods sector, food and beverage consumption, automotive sales, and iron and steel sectors. Demand forecasting methods such as artificial neural networks, machine learning, time series analysis, and statistical methods have been used in demand forecasting studies. By comparing the forecast performances of the used demand forecasting methods with each other, the forecast method with the best forecast performance was selected. Each of the studies contributed to the literature by using different inputs, different data, and different methods. In this study, the artificial neural network method, which is frequently used in the literature, was used. The artificial neural network was created inspired by the working structure of the human brain. It is generally used to make predictions, detect complex data patterns, analyze data, and voice recognition. In order to carry out the study's artificial neural network application, the criteria affecting the demand for the part were determined by first obtaining expert opinions and conducting a literature review. The determined criteria are truck production quantity, electricity unit price, Euro exchange rate, domestic producer price index, general consumer price index, Brent oil barrel price, scrap prices, peak prices, and gross domestic product. The output of the artificial neural network is the total sales units of the part family group. The data for the determined criteria are monthly data between 2017 and 2022. Data was collected from various sources to create inputs within the scope of the study. Before taking action on the data, it is necessary to analyze the data. Because there is a possibility of discrete data among the data, having such data in the data set reduces the study's accuracy. In order to analyze with appropriate data, outlier data analysis was performed with Box-Plot analysis, which is a data separation method, and outlier data were removed from the data set. After ensuring the reliability of the data, the data was normalized. The artificial neural network application created models with tansig and logsig activation functions. The prediction performance of both tansig and logsig activation functions was compared. The data that had to be defined in MATLAB was normalized to compare the prediction performances of these two activation functions. While the data was defined in MATLAB, no separate data was divided into learning, testing, and verification. The study was conducted using the random selection feature of MATLAB. MATLAB software separated the data as training, testing, and validation at 70%-15%-15%. After the data was defined, models with different layers and neuron numbers were created, including eight different learning algorithms: trainlm, traincgf, traincgb, trainingdx, traincgp, trainscg, trainingda, trainingd, which are included in the feedforward backpropagation algorithm, which is different from other studies in the literature. Predictive values were obtained by training the created models in MATLAB. In the literature, the prediction performances of artificial neural network models have been compared with mean absolute error (MAE), R Square, root mean square error (RMSE), and mean absolute percentage error (MAPE) analyses. In the study, analysis was made for all error analysis methods used in the literature. However, in the study, only the three models with the best prediction performance of the error analysis methods were shown. To determine the best artificial neural network model, based on the R Square error analysis method, one of the error analysis methods, it was concluded that it was an artificial neural network model with a traincgb learning algorithm, two layers, ten neurons, tansig activation function. After the artificial neural network application, ANN's forecast performance was compared with other demand forecasting methods used in the literature. For this, the ARIMA, a time series analysis method, was applied. ARIMA method is a statistical method used in time series analysis. It consists of AR-I-MA components. AR stands for autoregression, I stands for integration, and MA stands for moving average. These three components are brought together and used to predict the future. Parameters are determined by statistical methods to minimize error. The ARIMA method is used in many areas, such as weather forecasts, market research forecasts, and epidemic analyses. There are advantages and disadvantages to using the ARIMA method. It has many uses, such as economy, finance, and health. When the components are well adjusted, it provides the advantages of forecasting competence in short and medium-term forecasts, flexibility in different models, and accurate evaluation of the results since it is based on statistical theories. When there is not enough data, disadvantages arise, such as being unable to make accurate predictions at low frequencies, making complex-level predictions, and requiring expertise to determine the correct parameters. Since ARIMA is applied to stationary series, it is necessary to check whether the sales data are stationary. Stationarity means that there are no sudden increases or decreases in the series. It is understood from the time series, ACF, and PACF graphs that the series is stationary. At the same time, the Augmented Dickey-Fuller test is performed to check the stationarity of the series. Sales data are transferred to MATLAB, and data are defined for the ARIMA method through the econometric models toolbox in MATLAB. The stationarity of the series is checked by drawing time series, ACF, and PACF graphs. It can be seen from the graphs that the series is not stationary. However, the Augmented Dickey-Fuller test is performed to prove that it is not numerically stationary. As a result of the test, when the test statistic is greater than the critical value, it is clear that the series is not stationary. A difference operation is performed to bring the non-stationary series to stationary form. It was determined as I=1 by differencing. It is necessary to check the stationarity of the different series. The time series graph, ACF graph, and PACF graph of the differenced series were drawn in MATLAB. It can be seen from the graphs that the series has become stationary. However, the Augmented Dickey-Fuller test is used to prove stationarity numerically. According to the test result, it was determined that the test statistic was less than the critical value. With this result, it has been confirmed that the series has become stationary, and the AR-I-MA model has I = 1. After the I value was finalized as 1, different models were proposed for AR and MA values. Among the proposed models, the value that controls the quality and complexity of the model by minimizing the error is the Akaike Information Criterion value. According to the Akaike Information Criterion value, the ARIMA (1,1,1) model had the highest prediction performance among other ARIMA models. R Square error analysis method was used to compare the prediction performances of the ARIMA model with the highest prediction performance and the artificial neural network model. The artificial neural network model with a traincgb learning algorithm, two layers, ten neurons, tansig activation function, which was determined to have the highest prediction performance by the artificial neural network method, was compared with the ARIMA (1,1,1) model using R Square analysis. As a result of the comparison, it was concluded that the prediction performance of the artificial neural network model is better than the ARIMA (1,1,1) model. When the estimated sales values of the artificial neural network model and the actual sales values are compared on the graph, it is seen that the prediction performance is very high. A new demand forecasting method was developed for the business in response to the developing automotive industry, and a comparative demand forecasting analysis study was conducted for the foundry industry, which is not available in the literature. This study can be applied to different parts of the business to determine accurate demand forecasting methods. In addition, different data can be developed and expanded with different layers, neurons, activation functions, and learning algorithms to be used in capacity adjustments.