Conv-Attention Model Based on Multivariate Time Series Prediction: The Cyanobacteria Bloom Case

Abstract

Multivariate time series forecasting problems are an important part of research in various fields at all times, such as financial and stock markets, natural disasters, disease prevention. However, forecasting has always been difficult due to its own reasons or external factors. In this paper, we propose a brand-new Conv-Attention network (CANet) for harmful algal blooms prediction. To capture more spatial dimension feature information, the network extracts the context dependency from each time series, and at the same time obtains the impact score between the interacting time series. In the previous stage of training, the feature factors are acquired through different convolution kernels. Then attention mechanism is adopted to model the processes that depend on mutual influence. To further enhance the robustness of the network, the CANet incorporates simple MLP layer-assisted training. The experimental results show that our proposed network performs well under the evaluation of the performance index.