Smart Climate Hydropower Tool: artificial intelligence for effective
hydropower production forecast and management
Abstract
Smart Climate Hydropower Tool is an innovative web-cloud-based service
that implements a set of data-driven methods for river discharge. An
application for two catchments in South America is discussed (test
cases), where management of hydropower plants can benefit from knowledge
of incoming discharge forecasts up to 6 months in advance. SCHT has been
developed inside H2020 project “CLARA - Climate forecast enabled
knowledge service” and exploits several Artificial Intelligence
algorithms, evolving by R&D activity to test new available ones.
Although tangible results using AI have been published (i.e. Callegari,
et al., 2015, De Gregorio et. al 2017) challenges remain for seasonal
lead times and rainfall dominated catchments, where forecast of
meteorological variables plays a critical role. In this contribution we
show results of application of different AI algorithms (from supervised
learning regression techniques, to artificial neural networks). Each
algorithm is trained over past decades datasets of recorded data,
forecast performances are then evaluated using separate test sets with
reference to benchmarks (historical average of discharge values and
simpler multiparametric regressions). Major operative advantages of AI
with respect to mechanistic hydrological models include limited to none
a priori knowledge of involved physical phenomena, high level of
flexibility when managing heterogeneous sets of variables related to
discharge, and quick setup time of the forecast. Major efforts are
requested to identity informative input features ranging from earth
observation to gauging stations data, to public meteorological forecasts
(i.e Copernicus Climate Change Service-C3S). Using AI techniques many
combinations of features can be tested together, to predict river
discharge to the reservoirs, choosing the best performing one and
tailoring the service to the catchment of interest. Once trained, each
algorithm just needs to retrieve online data to perform forecasts, with
limited maintenance (i.e. annual re-training to consider new available
hydrological data). For demonstrational purposes we prototyped a
cloud-based service, for immediate visualization, through a common
browser, of both past and forecasted data, and get on fly performance
metrics calculation of the forecasts.