Accurate load forecasting is the most crucial step in optimal power system planning. Electric load is distributed in both space and time. The feeder level load accumulates to produce sub-divisions, divisions, circles, DISCOS, and then the National level load. Similarly, in the temporal dimension, the hourly load is accumulated to produce daily, weekly, monthly, and yearly loads. In the future, as various new types of loads (electric vehicles, net metering etc.) will be part of daily life, this division of electric load will become more dynamic.

Currently, Pakistan's power system has a simple trend-based forecasting approach. The percentage increase in load is added in forecasted years. The load is predicted at a sub-division and yearly level, which is then aggregated into different spatial hierarchies. For example: the forecasted sub-division load is aggregated into divisions, circles, DISCOs up to the national level. This simple trend-based approach is obsolete and produces inaccurate results causing substantial financial costs to the country.

According to the National Transmission and Dispatch Company (NTDC), the post-2020 expansion of power generation will raise the share of renewable energy in the system to 30% energy mix. However, due to the variable nature of renewable energy, the addition of these resources will produce variabilities and uncertainties in the power system. To ameliorate these effects, DSM programs will be an essential part of the modern grid. Most of these DSM programs work at an hourly or daily level and are guided by load forecasts. While testing of these DSM programs should be carried out for years for the feasibility analysis. The optimal design and analysis of these DSM programs require an hourly load forecast up to the year. Most of the renewable energy resources are distributed in space and smaller in magnitude. To cost-effectively integrate these resources, one needs a decentralized planning approach that is capable of dealing with the intrinsic uncertainties of renewable resources.

Certainly, the future of the Pakistan grid requires a new type of forecast at a finer spatial and temporal resolution. We propose a hierarchical load forecasting structure called Spatio-Temporal load forecasting. In which we divide space and time to use a bottom-up forecasting approach. We plan to forecast the hourly electric load for every feeder of Pakistan up to one year ahead (8760 hours). These feeder level forecasts can be aggregated to produce forecasts of different spatial (circles, DISCOs) or administrative units (districts, provinces). There are significant challenges in enrolling Spatio-Temporal load forecasting. One of the main challenges is the large horizon of forecast (future 8760 hours). This is like merging both short-term and mid-term load forecasting. To solve this problem, we propose a deep neural network architecture along with transfer learning. Spatio-Temporal load forecasting will lead to accurate load forecasting, a better understanding of the load at spatial and temporal level, optimal integration of renewable resources and electric vehicles in the grid.