Aplicando rede neural para realizar predições de níveis do Rio Tapajós na Região Oeste do Pará

Abstract

The objective of this study was to use an LSTM neural network to predict the water level of the Tapajós River, based on hydroclimatological data from the National Water Agency (ANA). Data from the years 1999 to 2023 were used from three hydro-telemetry stations strategically positioned along the river, providing hydrological information. The research adopted the Knowledge Discovery in Databases (KDD) methodology to assist in data processing and organization, aiming to improve the results and interpretations of the analyses. The LSTM model was built and trained to generate monthly predictions of water levels. Min-Max normalization was applied to standardize the input data, facilitating model convergence during validation and prediction processes. The model was configured to use a 30-day observation window. The MAE, MSE, and RMSE metrics were used to evaluate the model's performance; as a result, the LSTM model showed satisfactory performance in predicting the Tapajós River's water levels. The evaluation metrics indicated consistent values; considering the best results from the three datasets, the mean absolute error (MAE) was 0.035, the mean squared error (MSE) reached 0.002, and the root mean squared error (RMSE) was 0.046, in a time series containing 1,333 samples, indicating good predictive accuracy. The LSTM was able to capture seasonal trends and variations present in historical data. Based on this, an autoregressive prediction technique was used for the first five months of 2024 to generate forecast data and compare them with actual data available for this period.