Automated model construction for combined sewer overflow (CSO) prediction based on efficient LASSO algorithm

摘要

The prediction of combined sewer overflow (CSO)udoperation in urban environments presents a challenging taskudfor water utilities. The operation of CSOs (most often in heavyudrainfall conditions) prevents houses and businesses from flooding.udHowever, sometimes, CSOs do not operate as they should,udpotentially bringing environmental pollution risks. Therefore,udCSOs should be appropriately managed by water utilities, highlightingudthe need for adapted decision support systems. Thisudpaper proposes an automated CSO predictive model constructionudmethodology using field monitoring data, as a substituteudfor the commonly established hydrological-hydraulic modellingudapproach for time-series prediction of CSO statuses. It is audsystematic methodology factoring in all monitored field variablesudto construct time-series dependencies for CSO statuses. Theudmodel construction process is largely automated with little humanudintervention, and the pertinent variables together with theirudassociated time lags for every CSO are holistically and automaticallyudgenerated. A fast LASSO (Least Absolute Shrinkage andudSelection Operator) solution generating scheme is proposed toudexpedite the model construction process, where matrix inversionsudare effectively eliminated. The whole algorithm works in audstepwise manner, invoking either an incremental or decrementaludmovement for including or excluding one model regressor into,udor from, the predictive model at every step. The computationaludcomplexity is thereby analysed with the pseudo code provided.udActual experimental results from both single-step ahead (i.e., 15udmins) and multi-step ahead predictions are finally produced andudanalysed on a UK pilot area with various types of monitoring dataudmade available, demonstrating the efficiency and effectiveness ofudthe proposed approach.