Stochastic modelling and time seriesOpen · a thesis to build on

Wind power ramps: forecasting the swings that stress a grid

Hisham Ihshaish · Research themes

Grid operators fear the swing, not the average: wind ramps forecast as continuous events.

Wind is now a first-order component of the world's electricity: installed capacity passed one terawatt in 2023 [1], and every gigawatt of it inherits the atmosphere's moods. The event that operators fear is not low wind but the ramp, a large change in a farm's output over minutes to hours, because reserves have to be scheduled against it and a missed ramp is paid for at the balancing marketa's worst prices. Forecasting ramps is a young, awkward problem with its own review literature [2], and it sits between three fields I care about: time-series modelling, the physics of turbulent driving (the noise theme), and machine learning that must answer an operational question rather than win a benchmark.

OneWhat makes it scientifically hard

The field's first difficulty is that a ramp is not a natural kind. The standard approach picks a threshold, forty per cent of capacity in an hour, say, and turns a continuous swing into a binary event, which throws away exactly the information an operator wants: how hard, how fast, and how certain. Working with Ignacio Deza and our student Russell Sharp, we treated ramps with a non-binary ramp function instead, so the model learns the intensity of the swing rather than a yes or no [3, 4]. The second difficulty connects to the noise theme: gusts are not Gaussian, and the tail behaviour of the driving wind is precisely what decides the worst ramps.

TwoWhere we got to

The characterisation and forecasting study, at an operating wind farm in north-eastern France, was written for the climate-change workshop at NeurIPS 2021 [3], and Russell's Masters thesis carried the non-binary formulation through deep-learning forecasters and presented it at SEEDS [4]. The thesis is on this site in full (PDF, with a summary), and it is a real starting point: the formulation works, and nobody has pushed it as far as it goes.

ThreeData you can start with today

Start here This theme needs no data agreement. The NREL WIND Toolkit provides simulated power and meteorology for more than one hundred thousand sites [5], and the GEFCom2014 competition data give ten wind farms with hourly power and weather forecasts, together with a published account of what forecasting methods won and why [6]. A Masters project can be running within a week of starting.

FourWhere a project could go

  • Ramp-aware objectives: Train forecasters whose loss weights the swing by its operational cost, and compare against the binary-event literature on its own metrics as well as ours. MSc
  • Characterisation before classification: Fit the distribution of ramp intensities per site and season, connect its tails to the non-Gaussian noise machinery of the noise theme, and ask which sites are predictable at all. MSc or PhD
  • Transfer across farms: A forecaster trained on one farm meets another farm's terrain and turbines: what carries, what breaks, and can the ramp function be made site-invariant? PhD

Notes

a Reserves are generation held ready to cover sudden shortfalls; the balancing market is where a grid buys that cover at short notice, at a premium. ↩

References

  1. Global Wind Energy Council (2024). Global Wind Report 2024. Cumulative installed wind capacity reached 1,021 GW at the end of 2023. gwec.net/reports/globalwindreport
  2. Gallego-Castillo, C., Cuerva-Tejero, A. and Lopez-Garcia, O. (2015). A review on the recent history of wind power ramp forecasting. Renewable and Sustainable Energy Reviews, 52, 1148-1157. doi:10.1016/j.rser.2015.07.154
  3. Sharp, R., Ihshaish, H. and Deza, J. I. (2021). Wind power ramp characterisation and forecasting using numerical weather prediction and machine learning models. Preprint, SSRN 3997702, written for the Tackling Climate Change with Machine Learning workshop at NeurIPS 2021.
  4. Sharp, R., Ihshaish, H. and Deza, J. I. (2022). Integrating wind variability to modelling wind-ramp events using a non-binary ramp function and deep learning models. SEEDS 2022, International Conference for Sustainable Ecological Engineering Design for Society. Thesis: ihshaish.github.io/post/sharp
  5. Draxl, C., Clifton, A., Hodge, B.-M. and McCaa, J. (2015). The Wind Integration National Dataset (WIND) Toolkit. Applied Energy, 151, 355-366. doi:10.1016/j.apenergy.2015.03.121
  6. Hong, T., Pinson, P., Fan, S., Zareipour, H., Troccoli, A. and Hyndman, R. J. (2016). Probabilistic energy forecasting: Global Energy Forecasting Competition 2014 and beyond. International Journal of Forecasting, 32(3), 896-913. doi:10.1016/j.ijforecast.2016.02.001
← previousNetworks from large time-series collectionsAll themesnext →Adaptive learning and AI in education