AVI-Net: Audio-visual-integration inspired deep network with application to short-term air temperature forecasting

Accurate forecasting of air temperature plays an important role in epidemic disease prevention, agricultural safety, tourism management, etc., but is hard to accurately perform since the series as a natural process involves lots of stochastic and nonlinear fluctuations. Meanwhile, most existing deep forecasting networks attempt to improve accuracy by integrating some design tricks and lack reasonable guidelines and domain knowledge. Biologically-inspired ideas are promising to address the above drawbacks, and this paper develops an audio-visual-integration inspired deep network, namely AVI-Net, for short-term air temperature forecasting, including the feature capture (imitating the auditory canal and ocular globe), feature analysis (imitating the left and right cerebral hemispheres) and forecasting realization blocks (imitating the high-level brain region). First, the domain knowledge is injected in the AVI-Net from the selection-addition feature and the loss function constructed via the Gaussian kernel, improving the extraction capability of long-term trends and short-term non-linearities. Second, multiple Monte Carlo dropout layers are integrated into the AVI-Net to introduce the model uncertainty, obtaining confidence intervals and enhancing the application values. Finally, the AVI-Net deeply mimics the audio-visual integrated system in the flowchart, structure, and function, inheriting its superior analysis capabilities and having somewhat interpretability in the network design. Seven experiments and six discussions under two real-world datasets present that the AVI-Net has better forecasting accuracy and stability than 14 baselines, and is suitable to be the intelligent and reliable decision support system for related sectors.