Abstract:
Pakistan is facing electricity crisis for decades and Government of Pakistan (GOP) is taking measures by adding more power plants and integrating renewable generation into the grid. However, demand increase in electricity with pace is because of addition of modern digital devices and global warming. Losses in the transmission and distribution lines are consistently increasing owing to weak and old power infrastructure of Pakistan. Moreover, replacement and development of power infrastructure is a lengthy process by reason of high cost of equipment. In order to achieve supply-demand balance, power polices were drafted in the past to increase generation. However, unfortunately desired target was never achieved. Financial debacles and economic challenges are results of poor system planning and short-term policy measures clearly facilitating investors and political influencers. Existing scheduling of IPPs is based on long term flawed contracts which avoids fair merit order, perform incorrect forecasting, distribute unequal payments, set high per unit price, and interference of politicians. Solution to permanently eradicate energy poverty is fair, long term and unbiased energy policy. Policy recommendations proposed in this work are based on results of machine learning algorithms which encourage timely scheduling of IPPs and revision of power purchase agreements (PPAs).Machine Learning is used after anticipating GOP interest in reviewing energy policy and planning for restructuring in power markets at policy level. Existing time series models that are trained on the present data sets performs really well. Contrarily, Artificial Intelligence (AI) and Machine Learning (ML) approaches are independent in decision making. These approaches are useful for policy studies because they will save cost, increase independency, encourage feasible payment methods, introduce fair merit order mechanism, and support optimal long-term planning. Machine Learning algorithms used in this work as a proof of concept are Logistic Regression, K-Means clustering and Anomaly Detection with average accuracy greater than 70 percent.
