Power of tech tools: Andhra using AI to control electricity purchase costs

Day-ahead predictions are critical, as they have implications for how much electricity should be generated in coal plants and how much should be purchased through the open market

Last year, as the country began grappling with the Covid-19 pandemic, the Transmission Corporation of Andhra Pradesh (APTransco), the state’s electric power transmission company, adopted a new method to cut the growing cost of electricity in the state.
 
Power purchases from the open market account for 80 per cent of the cost burden of the power sector. Hence, accurately predicting the demand for power the next day is critical to controlling costs. For the past 20 years, APTransco had been relying on manual methods to estimate the demand for power the next day.
 
In this method, the load dispatch operators would estimate the demand depending on such factors as the season, the weather, and holidays. A lot depended on the current day's demand. “Suppose the demand was 9,000 megawatts at 6 pm on a particular day. It could be predicted to be of the same quantum the next day, with a maximum accuracy of 90 per cent,” explains Srikant Nagulapalli, Andhra Pradesh’s secretary for energy.

 

AP State Load Dispatch Centre (SLDC) officials knew that a digital model could enhance the accuracy of power consumption forecasts, thereby reducing the cost of power purchases from the open market and bringing down tariffs.
 
Forecasting electricity demand forecast is a big task for SLDCs, which need to prepare for the purchase or sale of power daily. A load-demand mismatch leads to over- or under-drawing from the national grid, which entails heavy penalties and sometimes leads to power cuts as well.
 
In Andhra Pradesh, the SLDCs decided to leverage technology to solve the problem. They began by making a list all the factors that would influence the energy consumption in the state, such as sunrise and sunset patterns, rainfall, and humidity.

 
Data on all these parameters for the last three years was collated, including demand scenarios during the lockdown and subsequent unlocking, to build a forecasting model based on artificial intelligence (AI) and machine learning (ML). Google’s open source platform, TensorFlow, was used to build the model, which would yield precise estimates of day-ahead energy demand in megawatts every 15 minutes. 
 
Day-ahead predictions become critical, as they have implications for how much electricity should be generated in coal plants and how much power should be purchased through the open market. The 24-hour forecast is also important, as the day is divided into time slots of 15 minutes each, with 96 time blocks.

 
If the forecast is accurately done for each block, as is done by this AI-ML model, then the exact amount of electricity required the next day can be dispatched every 15 minutes from each generating station, hence minimising power purchase costs and reducing wastage.
 
“If the forecast is not known, then our planning will go haywire and costs will increase,” explains Nagulapalli.
 
APTransco has come up with several versions of this model over the past year. The most advanced one is Dove 6, which has an accuracy of 97 per cent. As a result, the state, which consumes Rs 30,000 crore worth of power annually, is able to save anywhere between Rs 2 lakh and Rs 3 crore a day by buying and transmitting the right amount of electricity.

 
Work is also underway to enable the state to use this model for other day-ahead forecasts, including for wind energy, solar energy, and for predicting market prices, says Nagulapalli. 
 
Since 25 per cent of the state's electricity demand is met by renewable sources, a wind and solar energy forecast model becomes necessary. He also foresees this model being used for price forecasting, which plays an important role for power generators and distribution companies, as they rely on this information to develop their bidding strategies. If a power generator has an accurate forecast of the prices, it can develop a bidding strategy to maximise profit.

 
“We are also willing to partner with startups and companies which can commercialise this model,” says Nagulapalli. The company, he adds, will add more data in the future to fine-tune the model’s accuracy beyond 97 per cent.