應用類神經網路、卷積神經網路與長短期記憶模型預測動態隨機存取記憶體價格

Abstract

動態隨機存取記憶體(Dynamic Random Access Memory，DRAM)為最重要的半導體元件之一，隨著科技產品的日新月異，動態隨機存取記憶體之需求量逐年增加。近年來，由於動態隨機存取記憶體的產值佔全球半導體產業總產值之比重提高，比例已經趨近18%，預測DRAM的價格，對於記憶體製造商、資通訊廠商、與投資人之重要性不可言喻。預測DRAM價格，可作為記憶體製造商調節產能及定價之基礎，也可以作為資通訊廠商訂定零組件採購成本與投資人預測合理股價之依據。類神經網路的原理，為模擬人腦的學習過程，近年來，學者、專家嘗試提出各類新型神經網路模型，並已經成功應用於許多領域，預測準確率，甚至高於許多傳統方法。雖然動態隨機存取記憶體價格之預測非常重要，但少有學者使用新型神經網路，預測動態隨機存取記憶體價格。因此，本研究擬導入類神經網路、卷積神經網路與長短期記憶模型，預測動態隨機存取記憶體之價格，並且比較三種方法之預測準確率。 本研究將以近十年第三代雙倍資料傳輸率(Double Data Rate，DDR3)同步動態隨機存取記憶體(Synchronous DRAM，SDRAM)單位價格為基礎，實證研究三種方法於預測之準確率，並比較三種方法之預測準確率。依據實證研究結果，長短期記憶模型為較佳的預測方法，在周資料集與日資料集預測有最低的平均絕對百分比誤差(Mean Absolute Percentage Error， MAPE)，分別為4.187與3.961，準確率高於其他兩種方法。本研究結果指出，長短期記憶模型具有良好的預測能力，可以作為預測未來動態隨機存取記憶體價格的方法。

Dynamic random access memory (DRAM) is one of the most important semiconductor components. With the rapid development of technological products, the demand for DRAM is increasing year by year. With the continuous increase in production. In recent years, due to the increasing proportion of the value of DRAM to the total value of the global semiconductor industry, the proportion has approached 18%. Predicting the price of DRAM will be very important for memory manufacturers, information manufacturers, and investors. Correct prediction can be used as the basis for adjusting production capacity and pricing, and investors can predict a reasonable stock price.The principle of the neural network is to simulate the learning process of the human brain. In recent years, scholars have tried to propose various new neural network models, which have been successfully applied to many fields, and the prediction accuracy is higher than many traditional methods. Although the prediction of DRAM price is very important, in recent years, few scholars have used new neural networks to predict DRAM price. Therefore, this study intends to introduce an artificial neural network (ANN), convolutional neural network (CNN), and long short-term memory model (LSTM) to predict the price of DRAM, and compare the prediction accuracy of the three methods.This research will be based on the unit price of the third-generation Double Data Rate (DDR3) synchronous dynamic random access memory (SDRAM) in the past ten years, to empirically study the accuracy of the three methods in forecasting. In the empirical study of LSTM as a better prediction method, the weekly dataset and daily dataset have the lowest mean absolute percentage error (MAPE), 4.187 and 3.961, respectively. The accuracy rate is higher than the other two methods and can be used as a method to predict the price of DRAM in the future.