Seasonal streamflow drought forecasting based on pattern recognition concepts using statistical and machine learning approaches

Abstract

Understanding and forecasting drought events is crucial for effective water resource management and mitigation planning. Forecasting droughts is challenging due to their inherently complex patterns and dependencies. However, there is a tendency for droughts to occur during specific seasons or times of the year and exhibit distinct seasonal variability. This research focuses on analyzing seasonal drought patterns using a grouped data concept, where similar data points are aggregated into groups to represent distinct hydrological drought conditions. The objective is to develop a methodology that can effectively recognize and predict droughts based on these grouped streamflow data sets. In the proposed study exploratory data analysis techniques are used to recognize the seasonal patterns within the data to extract meaningful drought patterns from the streamflow data. The study employed a combination of statistical methods and machine learning techniques, including Markov models and Long Short-Term Memory models (LSTM), to forecast the grouped seasonal streamflow data. A Markov model is employed to model the transition probabilities among hydrological drought states, capturing the temporal dependencies in streamflow behaviour. Subsequently, a Hidden Markov model (HMM) is utilized to employ the underlying states (or underlying drought levels) in observed streamflow data. To further enhance forecasting capabilities, monthly and weekly LSTM networks are utilized to learn long-term sequential dependencies and forecast future streamflow drought patterns. The study area was selected as the Palliser Triangle, the driest region in Canada. A total of 25 river stations (catchment area ranging from 319 to 47,800 km2

) were chosen, representing a range of river capacities: low flow (annual runoff range from 0 to 50 mm), medium flow ((annual runoff range from 50 to 175 mm), and high flow (annual runoff more than 175 mm) The monthly flow sequences of these rivers displayed the coefficient of variation ranging from 0.61 to 3.84, skewness from 0.57 to 8.39 and lag-1 autocorrelation from 0.2 to 0.63. In view of the highly skewed nature of monthly flows, the Box-Cox transformation was applied to normalize the data sequences and the normalization parameter ƛ ranged from -0.96 to 0.16. The Box-Cox transformation proved powerful for the normalization of flow data sets, which provided a strong platform for the analysis and forecasting of hydrologic droughts. The model results revealed that the discrete Markov model performed best for medium-flow rivers, achieving an average forecast accuracy of 65%, and the Hidden Markov model demonstrated superior performance for both low-flow and high-flow rivers, with an average forecast accuracy of 74%. The LSTM model showed consistent performance across all river types, providing monthly forecasts with approximately 80% accuracy and weekly forecasts with an impressive 90% average accuracy. [...]