Shallow learning model for long-term cyanobacterial bloom forecasting in real-time monitoring system

The success of Deep Learning (DL) methods in recent years has popularized complex architectures with extensive layers and parameters, enabling models to capture intricate relationships and extract relevant hidden features. While these architectures achieve impressive results in many classical applications, they are often prone to overfitting and are too costly to be implemented for edge-computing applications, particularly in real-time series forecasting tasks. This paper introduces a shallow Long-Short Term Memory (LSTM) neural network model capable of forecasting cyanobacterial blooms with up to a 70 % accuracy for a 28-day time horizon. This model is embedded in a micro controller unit after applying a quantization process. Unlike traditional methods that rely on centralized processing, our edge-based approach offers real-time, on-site forecasting capabilities, reducing latency and dependency on external infrastructure. We propose it as a cost-effective, low-power and easy to implement edge-based AI system for monitoring buoys, capable of broadcasting predictions and raw measurements through wireless communication. The performance of our model is evaluated with respect to state-of-the-art models and results are obtained for four forecasting horizons (16, 20, 24, 28 days) using Mean Absolute Percentage Error (MAPE). It shows to be 10 points more accurate than the other considered models for the worst-case scenario. The proposed system can be used to aid human forecasting experts or as a standalone system.