Agricultural product price prediction based on signal decomposition and deep learning

Abstract: A stable price of agricultural products is of great significance to the social economy and agricultural development in recent years. But, it is difficult to accurately predict the agricultural product prices, due to the non-stationary, non-linear, and high volatility. In this study, a novel prediction model of the decomposition-reconstruction-extraction-associated-output agricultural product price (CT-BiSeq2seq) was proposed using signal decomposition and deep learning. The multi-dimensional data was added to improve the model prediction accuracy, such as the average temperature, and fertilizer cost (price of pig formula feed and urea). Firstly, the original price series were divided into simple ones using the complementary ensemble empirical mode decomposition (CEEMD). Secondly, the original price series was reconstructed into the high-frequency, low-frequency, and residual items, according to the Pearson correlation coefficients and the decomposed subsequence. Thirdly, the data features of the reconstructed sequence were extracted via a temporal convolutional network (TCN). The 7-dimensional data was input to extract the influencing factors on the price of agricultural products. The output steps were similar to the input ones. Fourthly, a Biseq2seq model was constructed with an encoder and a decoder. A bi-directional Long Short-Term Memory network (Bi-LSTM) was introduced into the encoder to strengthen the global correlation between sequence data. Finally, the LSTM network was introduced into the decoder to output the predictive value of the number of steps. Taking the pork price of the Fengtai District wholesale market in Beijing of China for empirical analysis, the prediction performance of the CT-BiSeq2seq model was remarkably better than the rest benchmark models, indicating the number of lags reached the optimal in 11 days. The mean square error (MSE), the mean absolute error (MAE), and the mean absolute percentage error (MAPE) were 0.657 4 rmb2/kg2、0.504 6 rmb/kg、2.116 7%, respectively. Furthermore, the few-day lag cannot fully reflect the overall characteristics of agricultural product prices, where there was easy access to fall into the local optimum. Once the lag days were too long, overfitting was easy to occur, leading to low prediction accuracy. An accurate and stable prediction was also achieved in other datasets. The MSEs of spinach, apple, and egg were 0.627 7 RMB2/kg2, 0.463 2 RMB2/kg2, and 0.552 6 RMB2/kg2, respectively, while the MAEs were 0.543 1 rmb/kg, 0.442 5 rmb/kg, and 0.533 9 rmb/kg, respectively, and the MAPEs were 3.204 7%, 2.236 1% and 2.231 4%, respectively. Therefore, the agricultural products with large price fluctuations were suitable for the large lag steps, whereas, the small price fluctuations were suitable for the small lag steps. A large number of lag days were completely learned from the trend in large price changes. The short lag days were used to fit the time sequence in the smaller price changes, due to the relatively stable trend of price change. Specifically, the prices of spinach and eggs fluctuated greatly in the data range, where the loss error reached the minimum over the 11 lag days, respectively. By contrast, the price of Apples fluctuated less over the 7 lag days. This model can provide a strong reference to forecast the price fluctuation of agricultural products.