Innovative Approaches for Mapping the Pressurized Irrigation Systems Performances Under Unsteady Flow Conditions

Nowadays, the management of pressurized irrigation networks requires plenty of information to provide an efficient and reliable service to farmers. An approach called MASSPRES is being developed as a collaboration between FAO and CIHEAM-Bari with the goal of developing a reliable modernization strategy and improving the performance of pressurized irrigation systems. Mapping the perturbation, which is represented by the unsteady state flow analysis, is one of the most significant steps of this approach. The perturbation in irrigation networks is often created when sudden changes in flow rates occur in the pipes. This is essentially due to the manipulation of hydrants (service outlets) according the operational scenarios called configurations. During the perturbation occurrence, pressure waves propagate through the networks pipes that may lead to a signification pressure variation. This variation could expose the irrigation system’s components to a substantial danger that could cause significant damage. To model such a phenomenon, several computational algorithms have been developed. The majority of these models aimed to simulate the unsteady state conditions induced by the farmer’s behavior. The most recent ones are efficient enough to provide a good image of the perturbation occurrence through different indicators, however, one of the main draw backs of such model is the significantly high time and computational costs. In the present work, two different generations of models were developed. The first is a directly programmed model that was devaloped based on the method of characteristics and two indicators have been introduced: i) The hydrant risk indicator (HRI), which is defined as the ratio between the participation probability of hydrant no. x in the riskiest configurations and its total number of participations; and ii) the relative pressure exceedance (RPE), which provides the variation of the unsteady state pressure with respect to the nominal pressure. The two indicators could help managers better understand the network behavior with respect to the perturbation by defining the riskiest hydrants and the potentially affected pipes. Although, knowing the riskiest hydrants in the network is an important piece of information, managing ramified networks in real time will remain a difficult task to handle in real time. Thus, the need of developing a real time Decision Support System (RTDSS) that could process such information and guide the manager in real time is crucial. For this aim, two thousand configurations (operational scenarios) were simulated using the directly programmed model from the first step and fed to train a new model based on deep learning with the objective of forecasting the maximum pressure occurred due to the perturbation at each section. The occurred pressure is represented as classes according to the case sensitivity and the required precision. Steps of 1, 2 and 3 bars were simulated. The model proved to be significantly time saving compared to previous approaches as the results are produced instantaneously with a forecasting accuracy of 85 %. Furthermore, using the confusion matrix, the error committed by the model is of one class lower or higher that may be considered tolerable according to the system sensitivity. This approach was applied on a pressurized on-demand irrigation system located in south of Italy that consists of 19 hydrants and covers 57 hectares. Nonetheless, the deep learning-based model needs to be trained on each section. Thus, as a main step of the method of characteristics, the network was discretized into 1017 sections of 3 meter each. Training the deep learning model for such number of sections is not practical and time consuming. For this reason, a code was developed using autoencoding combined with t-distributed stochastic neighbor embedding (t-SNE) algorithm for features extraction and their visualization respectively. It is principally to cluster the sections according to their behavior to the perturbation, thus, reduce the number of trainings of the previously mentioned model. Nine zones of similar behavior were determined by the present developed code and the deep learning model will be trained only on these zones representing all the sections. The two last developed codes could be integrated for a decision support system (DSS) for modelling the perturbation in the on-demand pressurized irrigation networks that would add a significant contribution to provide practical recommendations for real-time decision-making processes. which was not possible using directly programmed software.