Solar forecasting and its value
What a better solar forecast is worth to markets, system operators, and plant owners, and how to produce forecasts that quantify their own uncertainty.
Solar forecasts keep improving, but accuracy alone does not tell a system operator, a trader, or a regulator how much a forecast is worth, or how much to spend on a better one. My work asks what forecasts are worth in practice, what their errors cost, and how forecast uncertainty can be reduced and managed.
What I have worked on
- The value of forecasts and the cost of their errors. A review of what solar forecasts are worth across applications: bidding in electricity markets, power system operations, and household electricity bills. Most published analyses look at the transmission level, from either the market or the system operations side. Compared with the levelised cost of electricity of PV, the value of forecasts and the cost of their errors are considerable. The review sets out measures to reduce and adapt to solar uncertainty and variability, which system operators, market operators, and regulators can select according to their own circumstances.
- Forecasts that state their uncertainty. With colleagues at NUS, I worked on probabilistic forecasting, which gives a range with probabilities rather than a single value. The methods improved the reliability of one-hour-ahead probabilistic PV forecasts, and deeper neural networks with skip connections produced more accurate probabilistic load forecasts than earlier methods.
- Satellite data in place of ground measurements. Reliable ground measurements of irradiance are scarce and costly. Using 15 research-grade stations in Europe, South America, and Africa, forecasts built on bias-corrected satellite-derived irradiance were of comparable quality to those built on ground measurements.
- Using forecasts in operation. A dispatch strategy that uses load and irradiance forecasts for PV, battery, and diesel systems cut operating cost by 5% against two common benchmark strategies, while keeping frequency, voltage, and harmonics within limits on a real-time simulator.
Where I have applied it
From 2019 to 2022 I led the PV grid integration working group of Singapore's national solar forecasting programme for the Energy Market Authority (EMA) and Singapore Power (SP), which included analysing the value of forecasts and the cost of forecast errors. In Indonesia, I have trained PT PLN (Persero) staff on variable renewable energy forecasting through GIZ. More on my experience.
Questions I help answer
- What is a better solar forecast worth in this market or power system?
- Which forecast horizon and format (a single value or a probabilistic range) does an operator need for each decision?
- Where can satellite-derived data replace ground measurements, and where should a project invest in its own sensors?
- How should a hybrid or island system use forecasts to dispatch its diesel generators and batteries?
Publications
- 2024The value of solar forecasts and the cost of their errors: A reviewRenewable and Sustainable Energy Reviews 189, 113915 · doi:10.1016/j.rser.2023.113915
- 2021Novel forecast-based dispatch strategy optimization for PV hybrid systems in real timeEnergy 222, 119918 · doi:10.1016/j.energy.2021.119918
- 2020Improving probabilistic load forecasting using quantile regression NN with skip connectionsIEEE Transactions on Smart Grid 11(6), 5442–5450 · doi:10.1109/TSG.2020.2995777
- 2020Can we justify producing univariate machine-learning forecasts with satellite-derived solar irradiance?Applied Energy 259, 114122 · doi:10.1016/j.apenergy.2019.114122
- 2019Reliable photovoltaic generation forecasting via quantile determinationIEEE PES Innovative Smart Grid Technologies Conference (ISGT) · doi:10.1109/ISGT.2019.8791617
- 2017An ensemble machine learning based approach for constructing probabilistic PV generation forecastingIEEE PES Asia-Pacific Power and Energy Engineering Conference (APPEEC) · doi:10.1109/APPEEC.2017.8308947