Weather is an essential component of natural resources that affects agricultural production and plays a decisive role in deciding the type of agricultural production, planting structure, crop quality, etc. In field agriculture, medium- and long-term predictions of temperature and humidity are vital for guiding agricultural activities and improving crop yield and quality. However, existing intelligent models still have difficulties dealing with big weather data in predicting applications, such as striking a balance between prediction accuracy and learning efficiency. Therefore, a multi-head attention encoder-decoder neural network optimized via Bayesian inference strategy (BMAE-Net) is proposed herein to predict weather time series changes accurately. Firstly, we incorporate Bayesian inference into the gated recurrent unit to construct a Bayesian-gated recurrent units (Bayesian-GRU) module. Then, a multi-head attention mechanism is introduced to design the network structure of each Bayesian layer, improving the prediction applicability to time-length changes. Subsequently, an encoder-decoder framework with Bayesian hyperparameter optimization is designed to infer intrinsic relationships among big time-series data for high prediction accuracy. For example, the R-evaluation metrics for temperature prediction in the three locations are 0.9, 0.804, and 0.892, respectively, while the RMSE is reduced to 2.899, 3.011, and 1.476, as seen in Case 1 of the temperature data. Extensive experiments subsequently demonstrated that the proposed BMAE-Net has overperformed on three location weather datasets, which provides an effective solution for prediction applications in the smart agriculture system.

No abstract