关键词: 粮食产量, 多因素时间序列预测, 深度学习, 智能算法

Abstract: To meet the demand for domestic food and ensure its security, agriculture has always been regarded as one of China’s important strategic industries. An accurate prediction of grain production is helpful for adjusting agricultural production policies and controlling the national macro-economy to ensure food security. However, an accurate grain production forecast is highly challenging due to various factors such as natural, social, and technological factors, and agricultural production conditions.
To more accurately predict China’s total grain production, a grain production forecasting model based on the adaptive differential evolution algorithm (ADE) and attention-based Bi-IndRNN (bidirectional independent recurrent neural network) is constructed. This model considers five major influencing factors: grain crop unit yield, agricultural production conditions, technological factors, agricultural insurance, and market and economic factors. Using IndRNN for predicting China’s grain production has the following advantages: it addresses the issues of gradient vanishing and exploding by adjusting the temporal changes in gradient backpropagation; IndRNN performs well when using activation functions such as ReLU, ensuring robust training; multi-layered IndRNN can be effectively stacked, especially with residual connections, increasing the depth of the network; the behavior of each IndRNN neuron in every layer is easily interpretable as each neuron is independent. Therefore, employing IndRNN neural networks in deep learning models may achieve better accuracy and stability in grain production forecast. However, setting the parameters for IndRNN is complex, requiring the selection of hyperparameters such as the number of layers, time steps, unit per hidden layer, batch size, and learning rate. Finding the optimal combination of these five hyperparameters is a highly intricate problem that directly impacts prediction accuracy and stability. Hence, a reliable and efficient algorithm is needed to accomplish this task. ADE with adaptive mutation factors strikes a good balance between global search and local search, possessing strong global search and convergence capabilities. Therefore, this study employs ADE to search for the hyperparameters of IndRNN.
Data on total grain production in China from 1991 to 2019 are collected from the National Bureau of Statistics. Taking into account the analysis of influencing factors on grain production by existing scholars and the fact that different stages of agricultural development in China are influenced by different factors, the following factors are selected as explanatory variables to analyze the main factors affecting China’s total grain production at present. These factors are divided into five categories: grain crop unit yield (grain unit yield per area), agricultural production conditions (the number of agricultural, forestry, animal husbandry, and fishery workers, total sown area of grain crops, rural electricity consumption), technological factors (the irrigated area of cultivated land, total power of agricultural machinery, amount of chemical fertilizer used), agricultural insurance (agricultural insurance premium per capita, average compensation per capita in agricultural insurance), and market and economic factors (grain import volume, agricultural production input index, grain production price index).
The predicted grain production for China from 2020 to 2024 is 667 million tons, 672 million tons, 680 million tons, 699 million tons, and 702 million tons respectively, showing an overall upward trend with an average annual growth rate of 1.15%. Furthermore, through the analysis of attention weights on multiple variables, it is found that the three variables contributing the most to the prediction of China’s total grain production are: grain unit yield per area, total sown area of grain crops, and irrigated area of cultivated land. Additionally, government subsidies for agricultural insurance, grain import volume, grain production price index, and agricultural production input index also contribute to the increase in China’s total grain production, providing suggestions for the development of grain production in China.
Currently, the main focus of China’s efforts to increase grain production and ensure food security lies in three aspects: grain yield per unit area, agricultural land, and irrigation area. To improve grain yield per unit area, it is crucial to enhance agricultural technology and improve planting techniques, as well as provide the guidance for the use of agricultural machinery. In terms of agricultural land, the “bottom-line thinking” on the total sown area of grain crops should be upheld, and strict measures for land protection should be implemented to protect the 1.8 billion mu (approximately 120 million hectares) of arable land, preventing its conversion to non-agricultural use. Increasing the irrigated area of cultivated land requires greater investment in agricultural water conservancy, improving irrigation systems, and focusing on the renovation and construction of irrigation pump stations and supporting facilities.
Individual insurance premiums per capita and average compensation per capita in agricultural insurance have certain effects on predicting total grain production. This indicates that the government’s implementation of agricultural insurance subsidies has alleviated the negative impact of natural disasters on grain production, reduced agricultural development risks, safeguarded the interests of farmers, and enhanced their enthusiasm for crop cultivation, thereby temporarily increasing China’s total grain production. Therefore, China should actively promote policy-oriented agricultural insurance development to provide protection for farmers against natural disasters and ensure food security.
Grain import volume, grain production price index, and agricultural production input index also play a role in predicting China’s total grain production. China has consistently adhered to the work principle of being “mainly self-sufficient in grain with imports and exports as supplementary”. However, due to changes in China’s grain production and consumption structure since the reform and opening up, China’s grain import volume has been increasing, making it more reliant on the international grain market. This has to some extent affected China’s food security and led to the introduction of relevant policies in recent years to stimulate the domestic grain production. Agricultural production input index and grain production price index reflect changes in the cost of grain production and grain prices, directly affecting the enthusiasm for grain production in the following year, thus being closely related to grain production.

Key words: grain yield, multivariate time-series forecasting, deep learning, intelligence algorithm