Peningkatan Akurasi Algoritma Long Short Term Memory (LSTM) pada Studi Kasus Prediksi Harga Cabai

Abstract

Unstable chili prices are still a serious problem in society. Many previous studies have provided solutions by predicting chili prices using various algorithms, one of which is Long Short Term Memory (LSTM). However, so far, there have been no prediction results that are considered quite accurate and consistent. The main causes are gradient problems, overfitting, and high error values in LSTM. So, to overcome these problems, start with how to properly process missing value data, namely using interpolation. Next, LSTM was developed to predict chili prices. This development aims to improve the weaknesses of LSTM, which focus on improving the high error value, so that it can have an impact on more accurate and more consistent prediction results.Some stages of the LSTM process are the forget gate, input gate, output gate, cell state (ct), and hidden state (ht). At the cell state (ct) and hidden state (ht) stages, development is carried out with the aim of overcoming the problem of gradient, overfitting, and high error values so that better prediction results are obtained. After all the processes are carried out, the results of the accuracy of the error value in the system from the development of LSTM in Labuhanbatu district are MAE = 2.589, RMSE = 3.419, and MSE = 11,695,900. This value is lower than the process results using the original LSTM and only using dropna in processing missing value data, namely MAE = 5,517, RMSE = 7,930, and MSE = 62,900,289. Then, the percentage of error value reduction from the comparison is MAE = 53,07%, RMSE = 56.88%, and MSE = 81.41%. It is expected that the low error value results from the development of LSTM can be an indicator of the accuracy of chili price prediction results, make the results more consistent, and overcome public anxiety.