Seasonal River Discharge Forecasting Using Support Vector Regression: A Case Study in the Italian Alps

Mattia Callegari,Ludovica De Gregorio,Claudia Notarnicola,Alberto Pistocchi,Marcello Petitta,Luca Pasolli,Paolo Mazzoli

doi:10.3390/w7052494

Mattia Callegari, Ludovica De Gregorio + Show 5 more

Open Access

https://doi.org/10.3390/w7052494

Copy DOI

Abstract

In this contribution we analyze the performance of a monthly river discharge forecasting model with a Support Vector Regression (SVR) technique in a European alpine area. We considered as predictors the discharges of the antecedent months, snow-covered area (SCA), and meteorological and climatic variables for 14 catchments in South Tyrol (Northern Italy), as well as the long-term average discharge of the month of prediction, also regarded as a benchmark. Forecasts at a six-month lead time tend to perform no better than the benchmark, with an average 33% relative root mean square error (RMSE%) on test samples. However, at one month lead time, RMSE% was 22%, a non-negligible improvement over the benchmark; moreover, the SVR model reduces the frequency of higher errors associated with anomalous months. Predictions with a lead time of three months show an intermediate performance between those at one and six months lead time. Among the considered predictors, SCA alone reduces RMSE% to 6% and 5% compared to using monthly discharges only, for a lead time equal to one and three months, respectively, whereas meteorological parameters bring only minor improvements. The model also outperformed a simpler linear autoregressive model, and yielded the lowest volume error in forecasting with one month lead time, while at longer lead times the differences compared to the benchmarks are negligible. Our results suggest that although an SVR model may deliver better forecasts than its simpler linear alternatives, long lead-time hydrological forecasting in Alpine catchments remains a challenge. Catchment state variables may play a bigger role than catchment input variables; hence a focus on characterizing seasonal catchment storage—Rather than seasonal weather forecasting—Could be key for improving our predictive capacity.

Highlights

The prediction of water availability is a key element for effective water storage management [1]
We have presented an application of Support Vector Regression (SVR) models for predicting the monthly mean discharge over 14 catchments in the Alpine region of South Tyrol (Northern Italy) using monthly discharge and snow-covered area (SCA) of antecedent months as input features
The comparison between the SVR models and the long-term monthly mean discharge, as well as with a simpler linear autoregressive model, is encouraging, with a mean improvement of the RMSE% on the 14 catchments selected in this study of 11%, 5%, and 2% for a prediction lag equal to one, three, and six months, respectively (Table 2), and a reduction in frequency of larger errors (Figures 8 and 9)

Summary

Introduction

The prediction of water availability is a key element for effective water storage management [1]. The statistical prediction of a variable ( + ∆ ) at an instant ∆ from current time t may be framed as a particular case of regression problem, the goal of which is to identify a function ( ) of a vector. The variables p1, ..., pm forming vector x may include current and past data samples [ ( ), ( − 1), ... Traditional regression methods seek a function that minimizes prediction errors considering all N samples. The support vector regression (SVR) technique [30,31,32,33,34], instead, aims at finding the simplest function that can fit all the data while minimizing the sum of prediction errors above a predefined threshold. For the sake of a quick illustration of the technique, let us first refer to the linear case, where the function to be estimated takes the form:

Methods

Results

Conclusion