Using seasonal palaeo-flow reconstructions and artificial neural networks for daily water balance modelling: A case study from Tasmania, Australia

Robust hydroclimate risk assessment requires a thorough understanding of past climate variability, which can be achieved by supplementing short instrumental hydroclimate records with palaeoclimate data. However, long-term continuous simulation of catchments and storage modelling, essential for hydrological risk assessment, necessitates monthly or daily time series input data, while palaeoclimate records are typically available at annual or seasonal scales. Additionally, modelling operational water storages used for hydropower is complex, requiring inputs such as water extraction information, which are difficult to replicate due to their inherent variability. Based on Lake Burbury, part of Hydro Tasmania's hydroelectric scheme in southern Australia, we demonstrate a novel method through which seasonal flow reconstructions can be used for daily palaeo water balance modelling, coupled with an Artificial Neural Network (ANN) model to simulate storage extractions. We first developed two seasonal tree-ring based inflow reconstructions, an approximately 1000-year Austral summer and a 400-year Austral winter reconstruction. We then used these as a guide to bootstrap historical daily inflows and the ANN known as Long Short-Term Memory (LSTM) was trained to simulate extractions for hydroelectricity. A Source model of the Lake Burbury hydro-electric water supply system was prepared to simulate the daily surface water balance of Lake Burbury, including inflows, outflows and resultant storage levels over some 1000 years. The simulations were used to ‘stress test’ the current storage system under a broader range of climatic conditions than the instrumental period. Based on our simulation, a low flow period like that in the 18th century represents the highest risk to hydroelectricity production, while a repeat of 12-13th century conditions would be associated with the highest spill volumes and most reliable electricity production. Importantly, by extending the instrumental record, we can place contemporary trends in water availability in a longer historical context, better assess the likelihood of extreme events, and hence adjust plans to decrease the vulnerability of the hydroelectric sector (among other water users) to drought/shifts in climate. This approach requires collaboration between palaeoclimatologists, the modelling community, hydrologists and managers of natural resources and the built environment. • Provides the first high-resolution, millennium-long seasonal streamflow reconstructions for Lake Burbury, Western Tasmania. • Utilises Long Short-Term Memory (LSTM) artificial neural networks to accurately simulate hydroelectric extractions. • Develops and applies a daily water balance model to simulate water levels and operational spillway releases. • Highlights the importance of incorporating palaeoclimate data into modern water management strategies for future adaptation.