Reinforcement learning model for the operation of conjunctive use schemes

A Reinforcement Learning (RL) model is proposed to identify operational strategies for conjunctive use schemes. This model is based on neuron-like adaptive elements that learn on-line to avoid system failures. The strength of this approach is that the resulting operational strategy for the water supply scheme reflects the need to respond and adapt to discrete failure events. A second advantage of the RL methodology is the inherent ability of control elements to operate in a distributed manner. By responding to local state inputs combined with a combination of local and global performance signals, the individual control elements are capable of operating effectively with only a limited set of state variables. The implication of such localized control is the avoidance of the curse of dimensionality commonly exhibited in other methodologies. Application of the RL model is demonstrated using the Burncrooks reservoir complex in Scotland. The model learns to effectively avoid failures, resulting in improved operational reliability.