Introduction
In order to reduce CO2 emissions and as consequence of the rapid cost decline of wind and solar power, the share of renewable power production (especially wind and solar power) increase in many countries around the world. As renewable power production increases, the electricity systems face a new challenge due to the fluctuating and intermittent nature of renewable power production. As wind, solar, wave and hydro power all are highly depending on weather conditions, the ability to forecast the weather and accurately transform it to a power forecast, becomes as key competence. For countries with modest amounts of renewables, it is primarily an economic issue in order to keep balancing costs down (cost of standby capacity and cost of up and down regulating power plants). For countries like Denmark, where wind power alone makes up for more than 40% of the annual power consumption and where renewable power can make up for more than 100% of consumption in individual hours, accurate renewable power forecasting becomes system critical. For Brazil, the weather forecast focusing in the rainfall forecasting, with the goal of making the scenarios of inflows to the hydropower plants of the system is one the most important aspects of modeling and resolution of the problem of optimization of the daily operation of the Brazilian hydrothermal system. This methodology considers aspects related to the electric grid as well as aspects related to the operation of hydroelectric, thermoelectric and other components of the system (with more and more wind and solar plants). In order to determine the minimum cost operation of the system, by calculating an optimal order, targets are set for the reservoirs at the end of the horizon. The computational models for optimizing the dispatch of the system are NEWAVE (horizon of up to 5 years), DECOMP (up to 1 year) and DESSEM (2 to 3 days). The methodology for the formation of the Settlement Price of Differences (“Preço de Liquidação das Diferenças” – PLD), which is the short-term price of electricity in Brazil, is all based in the NEWAVE, DECOMP and DESSEM programs. Therefore with the increasing of wind and solar in the share of Brazilian Electrical Power Matrix, the weather (rain, wind and solar) and power forecasting is one of the most important issues to the Brazilian Electrical Sector.
The forecasting solutions provided by ENFOR has been operational since 1994 and have supported the transition of the electricity system both in Denmark and many other countries. The aim of the article is to share the experiences from this journey and thereby help electricity companies avoid some of the forecasting pitfalls which exists along the way to a sustainable and greener future.
Principles os Wind Energy Forecasting
Operational power forecasting typically has a time horizon of a few minutes ahead and up to approx. 2 weeks, which can thereby be used for short to medium term planning of power production. Such power forecasts are based weather input from global and regional weather models provided by players like the European Centre for Medium-Range Weather Forecasts (ECMWF), Global Forecast System (GFS) or national weather providers, as the “Centro de Previsão de Tempo e Estudos Climáticos” (INPE) from “Instituto Nacional de Pesquisas Espaciais” (CPTEC). The weather forecasts are then combined with local measurements from the renewable power assets as well as local weather measurements to produce power forecasts.
Accurate forecasts can be achieved by feeding off-line production data on a regular basis (once a month or similar) such that the advanced machine learning algorithms can automatically adapt to the data and thereby continuously tune the models for optimal performance even as the assets wear and tear. For increased short-term performance the access to online/real-time data from the assets are critical as huge improvements can be achieved on the 3-4 hour time horizon, if the forecast models are continuously calibrated with online production and availability data. It is also possible to combine off-line and on-line assets using upscaling techniques. In such a setup, the online data from some assets will benefit the forecasts of assets where only off-line data is available. Figure 1 shows that after 13 hours of forecasting horizon there is no difference between on-line and off-line data for forecasting.
Figure 1 – Mean Absolute Error – MAE (normalized with installed capacity of wind farms) as a function of the Forecast Horizon, showing the positive effect of online data on the short term forecast horizon.
With the development of the internet and cloud computing, the concept of “forecast as a service” has emerged and are becoming more and more popular due to its simplicity and effectiveness. Even though it is indeed still possible to setup a local installation in the clients Information Technology (IT) – environment, there are a number of benefits from letting the forecasting provider handle the installation. In such a setup the client just transfer measurement from the assets (e.g. production, availability and weather) to the forecast provider, who then produce the power forecasts and transfer them back. The provider will take of model updates, weather forecasts, server hosting (and redundancy), which often yields a more reliable service.
With 12 years of handling renewable energy forecasting all over the world, ENFOR has developed a long range of special forecasting techniques which have been imbedded into dedicated modules, which can handle special situations. With a background from the mathematical institute of the Technical University of Denmark (DTU), ENFOR has developed particular accurate probabilistic forecasting techniques such that the uncertainty of the forecast can be accurately described with quantiles and/or interval as showed in the Figure 2 below. Figure 3 shows the wind power forecasting, considering the forecast horizon (1 hour, 6 hours, 24 hours and 48 hours) in comparison with the measured wind power for a wind farms portfolio. These techniques can also be used to provide the probability of ramping events, for improved risk management.
Figure 2 – WindFor™ (ENFOR software solution for wind power forecasting). In the case, ensemble forecast module for generating fractiles and uncertainty bands, which can be used for optimizing trading strategies or operational risk assessment.
Figure 3 – WindFor™ – Wind Power Forecasting for 1 h, 6 h, 24 h and 48 h compared with Measured and Rated Power
Other modules have been driving by specific customer requirements as the Australian Energy Marked Operator (AEMO), for whom ENFOR, and its partner Overspeed, developed an ultra-reliant wind and solar power forecasting system with 100% uptime. The system has never had downtime during the 8 years of operation.
For Hydro-Quebec in Canada, who has a large wind power portfolio in mountainous and icy conditions, ENFOR developed special techniques for forecasting of wind power in complex terrain as well as forecasting icing event and the decay of such ice on the turbines.
More widely used, is the combination forecasting module, which can take multiple weather forecast providers as input and then automatically and dynamically select the best combination of weather forecast providers for the individual renewable power assets. By applying such techniques, it is often possible to further improve forecast accuracy by 5-10% – when comparing with the best stand-alone weather forecast provider, as showed in the Figure 4.
Figure 4 – Combining multiple forecasts based on different weather providers can typically reduce the forecast error by 5-10%
Data Requirements
When it comes to renewable energy forecasting, data is of uttermost importance. Obviously, there is a trade-off between the cost of implementing vast amounts of data processing and the value of the resulting improvements of accuracy. Complexity, and hence the cost of both setup and operation, increase with the amount of data. For smaller portfolios, the cost/benefit ratio of an advanced setup might not be high enough, and it might be advisable to go for a more simple but still fairly accuracy configuration. For large GW portfolios it is almost always worth the extra effort of including as much data as possible, since small improvements in forecast accuracy often have high economic benefits. The ENFOR solutions a flexible and are built in way such that they can provide reliable forecasts even with very limited data, but forecast accuracy improves, the more data there is available.
As described in the previous section online production measurements can have a huge impact on intra-day forecast accuracy. Schedules (e.g. outages and maintenance information) is another important but often overlooked piece of information, which surprisingly often is forgotten in a forecast setup. If an operator for example has planned maintenance on 50% of the turbines in a wind farm, but has not informed the forecast provider about it – then the production forecast will obviously be completely off. In addition, the omission of such information will not only impact the specific event, but also cause noise in the models and reduce the general forecast error. For optimal performance the forecast provider should not only automatically receive information about schedules, but ideally also receive online data about turbine availability (turbines ready to produce) and set-point (e.g. curtailments) if at all possible.
Forecasting of new wind and solar farms is an exercise, which requires particular attention as there exists no historical data. Ideally, the design of the wind or solar farm also resulted in a power curve describing the expected production as a function of the most important weather parameters. Such, a power curve is then used to initialize the forecast engine, which should then be tuned to quickly and fully automatically adapt the power curve as real measurements are recorded and feed to the forecast models. Alternatively, a default power curve based on turbines types and other relevant static information can be used to initialize the forecast process.
Selecting Forecast Provider
When selected forecast provider there is a number of items to consider, of which forecast accuracy, system reliability and price are probably the most important factors.
Even it might seem simple, forecast accuracy is actually a complicated subject, since it can be defined in so many ways, which makes it difficult to compare across forecast providers and across portfolios. In order to best compare between forecast providers, the performance metrics should be precisely defined and match the economics of the specific electricity market in question. Normalized Mean Absolute Error (NMAE) is one of the classic performance metrics used, which derives from the Mean Absolute Error (MAE). The MAE measures the average magnitude of the errors in a set of predictions, without considering their direction. It is the average over the test sample of the absolute differences between prediction and actual observation where all individual differences have equal weight.
The Normalized Mean Absolute Error (NMAE) normalizes the MAE by the dividing the capacity, which makes it possible to compare across different assets.
In markets where large errors have a disproportional cost compared to small errors, it is better to use the Root Mean Square Error (RMSE), as this metric penalize large errors more than small errors. The RMSE is a quadratic scoring rule that also measures the average magnitude of the error. It is the square root of the average of squared differences between prediction and actual observation.
For comparing RMSE across different wind farms or portfolio it is also normalized with the capacity and thereby results in the Normalized Root Mean Square Error. The similarities are that both NMAE and NRMSE express average model prediction error which can be compared across assets and portfolios. Both metrics can range from 0% to 100% and are indifferent to the direction of errors. Since they are both measuring the error, lower values are better. Figure 5 shows an example of NMAE evaluation for comparison of forecasting accuracy, considering different types of wind farm and the geographic regions, and comparing the results for the portfolio (total) against individual wind farm.
In addition, it is important to define the exact forecast horizon to be used (e.g. day-ahead or hour-ahead) as the forecast horizon should match the decision processes within the company and the regulatory framework in the country. Accuracy can only be compared if measured during the same period (and a minimum of 3-4 months) as there will be variations from months to months and from season to season.
Many other factors have huge influence on the performance, such as:
- The capacity factor of the farms;
- Geographic location as well the aggregation level used for reporting performance (a big portfolio have significantly lower forecast error than an individual farm;
- In particular, for solar farms it is important to exactly define how night hours are accounted for and excluded from the performance metrics (since including them will significantly decrease forecast error).
Figure 5 – Shows the monthly NMAE (for the forecast horizon of 13-37 hours ahead) for three different customers. As it can be seen, different types of wind farm and the geographic regions have big impact on the forecast performance. In addition, it is clear that forecast accuracy is significantly higher on the portfolio (total) level than on the individual farm level.
Forecast accuracy for a specific portfolio, can be measured through a real-time/online trial or through an off-line simulation based on historical data. Off-line simulations have the benefits that they are fairly quick and cost-effective for both parties. The big drawback is, that it is easy to tamper with the simulations and improve forecasts when both the actual production and actual weather is known, when doing the forecast. Therefore, simulations should only be used as benchmark when engaging with well-established forecast provider, where there has already been established a high level of trust. Real-time/online trials have the benefit that they reflect real performance, but they take time (at least 3-4months), are costly for both customer and forecast provider (in the long run customers will have to pay the costs of running trials even they might not face any direct costs when carrying out a trial). Figure 6 shows that the forecasting performance varies monthly, and it should be used at least 3 months of data for reliable wind energy forecasting.
Figure 6 – NMAE for a wind portfolio. An assessment of performance should at least use from 3 to 4 months of data, as performance varies monthly.
Since more complicated setups and special configurations can rarely be modelled during a real-time trial or off-line simulation, these benchmarks will only give an indication of the performance one can expect. Therefore, it is also important to check the experience, track record of the forecast provider and ideally receive references from customers, who should be well-known utilities.
The experience of the forecast provider to handle various special configurations, will in the end also have an influence on the outcome and forecast accuracy. In addition, the experience of the forecast provider will most likely also have a significant impact on the reliability of the forecast service and if they consistently will deliver forecast even when parts of the input data might be missing or contain errors. The ability to handle erroneous data and consistently deliver forecasts in such situations, comes with years of experience. Reliability should also be an evaluation criteria during a real-time trial, but most likely the best way to assess this factor, is to check references and the reputation of the forecast provider since trial often only run for (relatively) short periods of time.
Lastly, price will also be an important criteria for selecting the right forecast service. If at all possible the evaluation should be based on an assessment of the value of accurate forecasts. Depending on the market and regulatory framework, this could be measured against imbalance penalties from market mechanisms, regulatory performance requirements (penalties) or the cost of handling/correcting imbalances (securing and activating standby capacity). Well-established forecast providers should have differentiated services, such that the individual customer can get the service matching their specific needs and yield the optimal cost/performance ratio. Such optimal cost/performance ratio should be based on the mentioned economic assessment and then obviously take the size of the customer portfolio into consideration. Achieving a 1% performance improvement has a significantly higher value on a 4 GW portfolio than for a 50 MW park.
Conclusion
Wind forecasting requires a combination of experience and advanced algorithms in order to achieve a forecast setup, which is both accurate, reliable and cost effective. Understanding both weather forecasts and measurement from the wind farms, as well as how to configure the system in regards to specific customers needs, are all important factors. Evaluating the right forecast provider for a specific setup require clear guidelines and definitions of evaluation criteria, such that a potential benchmark of different providers yield the right result. Making the right data available for the forecast providers are critical in terms of achieving accurate forecasts and requires active participation from the utility company. Since wind power forecast is often important operational information, it is also important that the forecast system is reliable with a high availability. This require robust algorithms to handle erroneous and faulty data, which can include missing values, and the ability to produce forecast even under such conditions. Obviously, the cost of a forecast product plays an important role and the economic benefits of having and accurate forecast should be measured against the cost of the system to identify the optimal cost/performance ratio.
About the Authors and Companies:
(i) Mikkel Westenholz is managing director of ENFOR.
(ii) Sérgio Augusto Costa is founder and managing director of VILCO Energias Renováveis and EMD Brasil.
(iii) ENFOR (www.enfor.dk) is a company based in Denmark, established in 2006, it delivers software systems and services for operational energy forecasting and optimization, based on more than 25 years of research. It is leader in renewable energy foracasting solutions having client as Transmission Systema Operators, Energy Traders and Asset Owners in more than thirteen countries around the world us: AEMO (Australia); Hydro Quebec (Canada); E.ON (Multiple European countries); Vattenfall (Multiple European countries); DONG Energy (Multiple European countries); EirGrid (Ireland); LitGrid (Lithuania); HOPS (Croatia); Energinet.dk (Denmark), and others.
(iv) VILCO Energias Renováveis (www.vilco.com.br) is an Engineering and Consulting services company with focus in business and development of Renewable Energy Projects (Large Hydro Power, Small Hydro Power, Wind Power, Solar Power and Biomass), acting in the Brazilian market since 2011.
(v) EMD (www.emd.dk) is a global company supplying software and consultancy services for design, planning, documentation and operation of wind energy projects as well as complex distributed energy projects. EMD has more than 30 years of successful wind advisory around the world, +60 GW of onshore wind advisory and +6 GW of offshore. EMD Brasil (www.emdbrasil.com) is a joint venture between VILCO and EMD for providing services and softwares specific for Brazilian market since 2015. EMD Brasil provides the ENFOR renewable energy forecasting solutions with exclusivity for the Brazilian market.
