With the growing penetration of decentralised variable wind and solar power, the need for more and more accurate forecasting is growing. This is particularly important for the weather, which impacts on the generation output and in turn on the supply-demand balance.
When looking at this issue back in 2014, Vermont Electric Power Company (VELCO) found that there were no commercial forecasting packages that could meet its requirements. So, it set about the creation of its own.
Kerrick Johnson, VP of Strategy & Communications at VELCO, tells Engerati in an interview: “At the time renewable integration was a problem we saw coming – and I have to say we greatly underestimated just how fast the renewable growth would be.”
At the time VELCO was also starting to face a specific problem in transmitting wind energy from the wind farms in the north of the state to the population centres mainly in the south, with congestion issues emerging along with curtailment of the wind energy.
“Coupled with this we were facing more frequent and more extreme, and thus more costly weather events. Put together, the grid was clearly becoming more weather driven and the skills and tools to run it needed to evolve,” Johnson continues.
As an organisation used to dealing with data it was obvious that this would be key for a solution and so VELCO, unable to address the issue on its own, approached IBM’s research analytics team for assistance.
Hyperlocal weather forecasting
The outcome of that collaboration is a “suite of interlocking and inter-related models” based around a “hyperlocal, hyper accurate Vermont specific weather model.”
“Whereas typical weather models are set up at a continental or semi-continental scale, when we set up the model we were practically trying to track every gust of wind and every cloud,” says Chandu Visweswariah, CEO of Utopus Insights, which was spun out from IBM following the project.
“We are interested in specifics such as the wind speed at a 90m hub height and the three-dimensional direction of that wind and the solar irradiance at the PV panel location.”
The weather model feeds wind and solar farm energy generation models. The same weather model in turn feeds into a demand model, which also takes into account behind the meter photovoltaic generation. These two forecasts are then linked together in a stochastic engine in conjunction with a grid model.
This final model, which in practice is run twice per day, provides a 72h outlook in 10m intervals across the whole state of Vermont at 1kmx1km resolution.
Applying the weather model to all the solar and wind farms in Vermont, energy forecasting mean average errors are 6% for solar and 9% for wind, Visweswariah says.
Visweswariah says that the keys to the model are the curation of the data and the use of advanced data science and machine learning techniques to exploit that data.
In broad terms, how the modelling works is by starting with a hyperlocal forecasting system. This model is then used for hind-casting, or historical weather forecasting, and these historical forecasts are then used to ‘train’ the machine learning algorithms for the wind and solar generation and the demand.
At the time, those algorithms took about two years to develop and refine. However, with the process now completed and automated, similar models could be set up for almost anywhere in the world within a week, Visweswariah notes.
Johnson adds that the model also reflects the respective talents of the two teams, combining the know-how of utility operations with that of data analytics.
Managing renewables in Vermont
Johnson says that Vermont is among the top states in the US when it comes to renewable penetration. At peak on a summer day, renewables are contributing about 30% and the goal is to reach 90% for all energy use (i.e., including transportation and heating) by 2050.
“We know we are benefitting from the modelling,” he says. “For every 1% load reduction we can shave off the peak by better orchestrating resources, we save $1m in ratepayer dollars.
“We know also there are operational gains for us as a transmission utility, for example, in the short term load forecast.”
He adds that already in 2017, just three years after starting the project, VELCO has “company-wide KPIs specifically related to delivering on all the ways these tools can provide value by increasing reliability and lowering costs.”
Recommendations for TSOs
Johnson says that the challenges VELCO is facing with growing renewable penetration are common ones and that all companies with renewables on their grid will need some form of advanced weather and energy forecasting tools.
“Each TSO will need to decide what best suits them but ultimately there are fundamental skills and tools like these that every TSO is going to need to ensure system reliability and to lower the risk.”
He adds that as the renewables penetration goes higher, and includes greater proportions of newer resources such as electric vehicles and heat pumps, the optimisation or orchestration of those resources will become even more important alongside the forecasting.
“We have been fortunate to embrace and tackle the problem early but we still feel we are no more than scratching the surface so far.”