Forecasts are part of every business (or at least they should be?) management decisions and their accuracy plays a substantial role in the decision making process of a data driven executive. The emerging technologies and tools for data analytics have made it easier than ever for companies to employ forecasting methods in almost every aspect and company division. From sales metrics, to marketing, HR, financials and many others. They can all benefit from an accurate forecast of metrics.
Short term forecast in the photovoltaic industry
Before we dive into the reasons how energy forecasts in the photovoltaic industry can be used and why they are needed, I have to explain what short term forecasting means.
Short-term forecasting refers to energy yield (or other metric) forecasts for a period of less than a week ahead.
The reasons why a photovoltaic plant operator or owner need a forecast of energy yield depends on the business model they are employing and the type of the system. For self-consumption systems, it is to match the peak draw of energy to the peak energy generation curve, which leads to higher self-consumption and self-sufficiency rates.
For utility scale photovoltaic plants that are exporting energy to the grid, the reason is different. In the age where photovoltaic systems where not so common, it wasn’t employed much often. Photovoltaic systems are intermittent energy source which depend on highly dynamic weather conditions. When PV plants were providing only a small percentage of the grid’s electricity, it didn’t affect the grid to a level that would require planning in advance. Now that the energy from PV generates a significant portion of the energy in the grid, it is essential to have a forecast of how much it would be produced by PV to be able to plan for it.
For example, in the autumn or spring months where energy output from PV is highly variable due to intermittent clouds and changing weather conditions, a grid operator would need to plan for low energy periods and employ other energy sources to satisfy the energy requirement of the consumers. When the PV output is high, it can utilize other methods such as energy output limitation or energy exporting. Energy output limitation is costly for the PV plant owner, because his plant does not operate at full capacity and therefore he is losing potential profit. Energy exporting in most cases requires planning ahead, which again requires forecasting.
Many countries which have a substantial amount of PV in their energy portfolio, are placing PV operators on the electricity trading market to bid for day-ahead auctions or intra-day auctions. The bid constitutes the energy amount at a given hour and the price that they would sell the energy at. If they fail to deliver, they would be penalized (in most markets, but not everywhere, however more and more countries are moving towards penalization). If they under-bid the amount of electricity that can be produced by the plant, they are losing potential profit. So, they are incentivized to utilize advanced forecasting methods to be able to optimize for maximum profit.
The energy yield from PV plant depends on many factors, some of which are explained in my other article, with weather being the factor that is the most difficult to predict and poses the greatest uncertainty. Soiling of modules, shading, defects, availability and maintenance periods, efficiency and others being the rest.
NWP (numerical weather prediction) models together with energy yield simulation is most accurate for 1-7 days ahead. NWP models are being operated by research institutions and are run on super-computers because of the vast amount of data and processing required. They predict what the weather will be like in the future. They are updated on a daily or sub-daily basis and can have hourly resolution. The predicted weather data is then feed into a simulator which has information about the system specification and that will produce energy yield predictions.
Time-series prediction methods are more accurate for intraday predictions up to several hours ahead, most of the time up to 4-6 hours. These methods rely only on past energy yield data to predict about the future. Some models from this group include SARIMA, SARIMAX, LSTM neural networks, dynamic linear models and TBATS. Ensemble models can also be used, which means that several models are utilized and some method to determine one single output is being employed.
A combination of both methodologies can be utilized to minimize error and correct for bias that can be introduced by the simulator or the NWP model.
All of the techniques mentioned require extensive data collection, filtering, training and validation to obtain usable results. Small to medium utility scale plants can not afford the luxury of hiring data-scientists to create a bespoke model for their use only. So, to solve for that, Solar Data has created a solution for monitoring and forecasting that integrates with the plant’s infrastructure and uses a proprietary model that accurately forecasts energy yield. Contact us for a demo presentation or trial!