Applying Hybrid Models Based on Deep Machine Learning in Smart Agriculture (Case study: Forecasting the Future Price of Pistachios)

The agricultural sector is one of the most basic and vital component in the social and economic structures of any country. Today, with increasing in the world's population and needing to provide food on the other hand, and increasing in the price fluctuations of agricultural products on the other hand, traditional agriculture is no longer responsible for the sustainable food security of the world population. In recent years, the occurrence of two incidents of the spread of the corona virus and the outbreak of war in Ukraine, have made the price of agricultural products extremely unstable. Today, even many farmers and agricultural associations in developing countries are not aware of the changes in market prices and the latest technological developments in the field of agricultural product prices, and they do not have the ability to discover the optimal price for selling their products. In such a situation, the use of intelligent models in order to accurately forecast the price of agricultural goods is vitally important for farmers and agricultural sector activists.Smart agriculture is an emerging concept that involves the integration of advanced technologies to collect and analyze data in order to solve the challenges and problems of the agricultural sector. In the meantime, forecasting the price of agricultural products involves with some basic challenges; including: 1) Data of agricultural product price is mostly non-linear, unstable, non-normal, and noisy and follows chaotic behavior, 2) There is uncertainty in the forecasted data obtained from different models, 3) In the studies related to price forecasting, the "publishable base model" is not provided in order to provide the forecasted price values. Therefore, the aim of this study is to provide a non-linear hybrid intelligent model for accurate forecasting of the future price of pistachios in the field of smart agriculture through managing the multidimensional nature of data, considering uncertainty in the forecasting data and finally building a publishable base model in the field of product price prediction.The hybrid model proposed in this study has the following innovations; 1) the deep learning neural network model and the Auto-Encoder network have been used to forecast the agricultural product price and determine the optimal lag of price as an input variable simultaneously, 2) The Monte Carlo method has been used as a non-parametric method to provide a confidence interval and calculate the most likely price that can happen, 3) The practical application of price forecasting models, i.e., "publishable base model" is presented in order to provide forecasted price values.

The implementation of the proposed hybrid model in this study includes the steps of "data preparation", "data feature engineering", "training and testing the final deep learning neural network model", "building the optimal base model", "creating the most likely price scenarios" using the Monte Carlo method and "inferring new prices or making out-of-sample forecasting" with new data sets” by feeding new price data into the deep learning neural network model. In the proposed hybrid model, data mining techniques are used, including Wavelet Transform (WT), Long-Short Term Memory (LSTM), Auto-Encoder network (AE), Monte Carlo-Markov chain (MCMC) simulation method and the concept of "inferring new prices".In the data preparation stage, using methods such as data smoothing, data rebuilding, correction of duplicate data in several consecutive days, and correction of missing data, the continuous set of pistachio future price time series is prepared to enter the primary model. Also, the wavelet transform function has been used for de-noising the data, the Auto-Encoder network has been used to determine the optimal lag, the Monte Carlo-Markov chain simulation has been used to create the most probable price scenarios, and the deployment concept has been used for out-of-sample forecasting with new data sets. The data used in this study is the time series of the daily price of pistachio futures on the Iran Commodity Exchange in the period from 10/13/2019 to 12/14/2021 in Rials per kilogram.

The results of this study showed that 1) by using the wavelet theory to de-noise the data, the error rate of the price data was reduced and the data had a stable trend, 2) the results of the implementation of the Auto-Encoder network showed that the optimal lag of one can be used as an input variable to forecast the future price of pistachios, 3) The outcomes derived from employing Monte Carlo-Markov chain simulation, coupled with out-of-sample forecasting using the new dataset, reveal compelling insights into the future pricing of pistachios on the Iranian Commodity Exchange. According to the analysis, the most probable and sanguine projection places the future price at the price ceiling of 213 thousand Tomans. Impressively, the forecasted price exhibits a minimal variance of merely 0.7% from the actual observed price, attesting to the precision of the proposed model. The overall accuracy of the model stands commendably high at approximately 93%.

Based on the results, firstly, the forecasted price has a small error with the actual price and this small error shows the power of the built model in forecasting the future price trend of pistachios. Secondly, the alignment of the price resulting from the Monte Carlo simulation with the new price can also be used as a confidence index in risk management for traders and market participants. Thirdly, the process set is the most complete value chain in the production of price forecasting models. Therefore, the use of the proposed hybrid model and the use of the components used in it, i.e. wavelet transform function, Auto-Encoder network, deep learning neural network, Monte Carlo simulation and the concept of inferring new prices; are suggested.