Optimal on-farm irrigation scheduling with a seasonal water limit using simulated annealing

As water resources are limited and the demand for agricultural products increases, it becomes increasingly important to use irrigation water optimally. At a farm scale, farmer's have a particularly strong incentive to optimize their irrigation water use when the volume of water available over a season is production limiting. In this situation, a farmer's goal is to maximize farm profit, by adjusting when and where irrigation water is used. However, making the very best decisions about when and where to irrigate is not easy, since these daily decisions require consideration of the entire remaining irrigation season. Future rainfall uncertainty further complicates decisions on when and which crops should be subjected to water stress. This paper presents an innovative on-farm irrigation scheduling decision support method called the Canterbury irrigation scheduler (CIS) that is suitable when seasonal water availability is limited. Previous optimal scheduling methods generally use stochastic dynamic programming, which requires over-simplistic plant models, limiting their practical usefulness. The CIS method improves on previous methods because it accommodates realistic plant models. Future farm profit (the objective function) is calculated using a time-series simulation model of the farm. Different irrigation management strategies are tested using the farm simulation model. The irrigation strategies are defined by a set of decision variables, and the decision variables are optimized using simulated annealing. The result of this optimization is an irrigation strategy that maximizes the expected future farm profit. This process is repeated several times during the irrigation season using the CIS method, and the optimal irrigation strategy is modified and improved using updated climate and soil moisture information. The ability of the CIS method to produce near optimal decisions was demonstrated by a comparison to previous stochastic dynamic programming schedulers. A second case study shows the CIS method can incorporate more realistic farm models than is possible when using stochastic dynamic programming. This case study used the FarmWi$e/APSIM model developed by CSIRO, Australia. Results show that when seasonal water limit is the primary constraint on water availability, the CIS could increase pasture yield revenue in Canterbury (New Zealand) in the order of 10%, compared with scheduling irrigation using current state of the art scheduling practice.