Neural Network Forecasting of River Levels for Flood Prediction Across Forecast Horizons

The research examines river level prediction systems which serve flood warning systems to determine their advantages and disadvantages for river management. The document focuses on rising flood occurrences and intensities which scientists link to climate change while stressing the requirement for better flood prediction systems to reduce destruction. The discussion examines traditional physical and numerical models together with data-driven artificial intelligence methods which include Perceptron and LSTM and GRU neural networks. The research evaluates Tanaro River water level prediction in Italy through an analysis of hourly hydrological and pluviometric data to assess AI performance against established benchmarks. The research evaluates forecasting periods ranging from 1 to 24 hours to enhance the performance of flood warning systems and decision-making tools for flood risk management. In view of the above, the general objective is to develop and evaluate predictive models for river level forecasting using historical hydrological and pluviometric data from the Tanaro River basin in Italy, applying Perceptron, LSTM, and GRU neural networks, and comparing their performance with traditional benchmark models such as the Persistence Model for forecast horizons ranging from 1 to 24 hours ahead.This study is classified as applied, quantitative research with an experimental orientation focused on model creation, training, and evaluation for hydrological prediction.It uses statistical inference and machine learning techniques to evaluate historical water levels data in a bid to benchmark model performance under different conditions as well as forecast major events such as flooding.The research uses an empirical method because it applies real data sets and benchmark assessments to prove the proposed solutions in actual operational settings. The advanced modeling techniques showed exceptional performance in real-world hydrological prediction through their accurate predictions and their ability to handle changing environmental conditions. Future research requires expanding the dataset through extreme event collection while enhancing model adaptability for non-stationary climate patterns and creating immediate decision-support systems to maximize practical outcomes.