Multilayer perceptron neural network for flow prediction.

Area of Science:

• Hydrology
• Artificial Intelligence
• Environmental Modeling

Background:

• Artificial neural networks (ANNs) are effective for modeling complex, non-linear hydrological processes.
• Previous studies demonstrate ANNs' capability in predicting rainfall-runoff, groundwater levels, and water quality.

Purpose of the Study:

• To assess the efficacy of ANNs in predicting average and maximum daily river flow.
• To forecast river flow 1-2 days ahead in a small forest headwater basin in northwestern Spain.

Main Methods:

• Utilized historical flow and meteorological data (precipitation, temperature, humidity, solar radiation, wind speed) from 2003-2008.
• Employed two ANN models: ANN(1) using disaggregated climate data and ANN(2) using aggregated evapotranspiration data.
• Trained models on data from 2003-2007 and validated with 2008 data.

Main Results:

• Both ANN models achieved a high goodness-of-fit (R² > 0.95) during training.
• Validation using 2008 data yielded correlation coefficients exceeding 0.85 for predicted vs. observed flow.
• ANNs demonstrated strong predictive performance for both average and maximum daily river flow.

Conclusions:

• Artificial neural networks are a reliable tool for short-term river flow prediction.
• The study confirms ANNs' capacity to model hydrological processes with high accuracy.
• Both disaggregated and aggregated climate data approaches proved effective for ANN-based flow forecasting.