Investigating the capability of long short-term memory input layers for sliding windowed data for enhancing water quality parameter prediction in small fishponds

Abstract

This research contribution is a comprehensive discussion of the Long Short-Term Memory (LSTM) capabilities investigation in handling sliding windowed data, especially for enhancing water quality parameter prediction in small fishpond. This research used a public dataset from an aquaponic system containing some information such as water pH, total dissolved solids (TDS), and water temperature. This research experiment on four different input types for our proposed predictive model using LSTM with the best handling way for processing sliding windowed data was input type 4, which used a flattening process with a timestep sequence order, resulting in the lowest error of 31.3912 in MAE and 87.3414 for predicting TDS, the lowest error of 0.1769 in MAE and 0.2375 in RMSE for predicting water temperature, and still good enough for predicting water pH with 0.6367 in MAE and 0.8302 in RMSE.