Wind and solar energy are two major components of renewable energy for a sustainable energy future of India. Increasing energy demand, non-availability of conventional energy resources and amount of emission of pollutants from commercial energy generation are major critical issues in considering wind and solar as alternative energy sources. The wind flow produces aerodynamic forces on the turbine blades to rotate and the rotation of the turbine’s shaft is transferred through a gearbox to an electrical power generator, which produces the power into the electrical grid system. Since the wind flow is a natural phenomenon, the power derived from the wind sources shows high variability and intermittency.
Though a small penetration variable energy like wind on the existing grid can be smoothly integrated; large penetration of wind energy can break the grid stability and hence the flexibility of grid operation is necessary. Forecasting and Scheduling (F&S) of wind power generation is an important factor to reduce the uncertainty of the wind energy generation such that the wind power variability can be precisely accommodated. A prediction of wind energy available for generation at specific time span in future is forecasting which is generally done with the help of different computational models combining meteorological data and plant characteristics. The generating schedule is a schedule for proper energy dispatch and is created using forecasting considering the issues of grid availability, maintenance schedule etc.
Due to variability and uncertainty, forecast- ing is an important aid to effective and efficient planning; the forecasting and scheduling strategy in variable renewable energy is an important factor in grid stability for high penetration of wind energy. Though the scheduling is mandatory with effect from January 1, 2012 in India, the Central Electricity Regulatory Commission (CERC) introduced a robust framework to strength- en forecasting in the renewable sector in India. The CERC also issued the Indian Electricity Grid Code (Third Amendment) Regulations, 2015 and Deviation Settlement Mechanism and related matters (Second Amendment), Regulations 2015.
Since blowing wind is a natural phenomenon, the wind power generation can be viewed as a complex real system and to predict the wind power generation one need to understand the physical theory and modelling of the complex systems which exists in time and space. A model simplifies the system-representation at particular point of time and space and the model based simulation manipulates the model parameters enabling to perceive the interactions and predict the different scenarios.
Wind power generation forecasting can be done using different models accommodating different observations like real-time or historical data. The most common way is using NWP (Numerical Weather Prediction) model in which different physical variable is simulated solving few differential equations representing the physical phenomena. This method is highly efficient in long term forecasting. The statistical approach in forecasting method is good for short term forecasting but sometimes it under performs the forecasting as it does not consider the physical behaviour of the parameters. At present machine learning and ANN (Artificial Neural Network) based system using NWP and statistical models became the well-accepted methodologies in wind power generation forecasting. But one must consider that all models are wrong, but some are useful, and F&S service providers provides the most useful models customized for a wind plant such that the model would produce the minimum error and maximum accuracy.
Though F&S of wind power generation is an important factor for sustainable energy future of India due to high penetration of wind power in existing grid, the forecasting technology opens some interesting issues in practice.
Use of Generalized model or customized forecast model
The common methodology to forecast is to create a generalized model and to change the parameter values depending on the plant characteristics. In most of the cases this methodology works but plant specific models considering the system equations of the plant sometimes shows better results. Not only considering plant specific parameters in generalized model, creating system representation of the wind plant without generalized system equations is another choice of forecasting methodology.
Aggregated forecast is better
It is a common belief that the aggregated forecast without plant specific forecast is better as it nullifies the positive and negative error in wind power generation. But the aggregated approach is a statistical methodology using some machine learning algorithms and hence sometimes aggregated forecast can not differentiate the root causes of error in aggregated fore- cast due to different plants. For example, if two different plants are connected in same pooling station and the aggregated forecast is only at the pooling station level, the proportion of penalty due
to deviation of two plants can not be calculated precisely if one plant is under performing. The aggregated forecast without using generator level forecast can create other issues like choice of grid resolution in computational architecture of NWP models, proper mix of different models and availability of useful data in different locations.
Multiple revisions are better
Multiple revision in wind power generation forecasting is required but upto a certain levels. Multiple revisions some times over shadow the actual behav- iour of the physical parameters and can over (or under) estimates the forecasting which can increases the total penalty due to deviation even in a day. Moreover the multiple revisions can break the dynamic stability of the forecasting as each time a forceful penetration of new patterns due to revision can create false scenarios which can increases the forecast error.
Forecast needs a lot of data
A forecast becomes good when in considers different patterns but does not replicate similar patterns. Forecast depends on the data availability, but the number of inputs depends on the plant characteristics, quality of the available data and perhaps some creativity. People believe those input have some predictive power over the forecast power. A proper choice of input parameter depending on plant characteristics is required to determine the forecast systematically. The correlation, cross correlation and principle components or Eigen vectors are not all time representation to define potential independent variables. Hence using of data intelligently rather than using all available data creates an optimized solution of forecasting.
A fixed time frame is required for learning of forecast model
It is a common belief that forecast is required for few months to stabilise the forecast system such that the system is able to self-learn using their feedback or feed forward rules. But the forecast systems do not always require a lot of data and the system must be retrained regularly. A good forecasting system can creates a stability problem if the same system runs considering very old data. The forecast system is much more dependent on its architecture and a good forecasting system follows supervised learning methodology. Hence creating a stable system is good for fore- casting but it is not good to stabilize the forecasting system using input data.
Forecasting accuracy is a number
Defining accuracy in forecasting is problem specific. The accuracy can be defined using RMSE(root-mean-square-error), MAE (mean absolute error) or simple can be defined on the penalty due to deviations. The accuracy of forecast depends on the data availability, quality of data and plant characteristics; and obviously on the choice of models. Same accuracy can not be achieved in different plants at different locations in different sessions. The forecast accuracy is not a simple number but a range of values showing the maximum and minimum accuracy in wind power generation forecasting.
Forecast value is deterministic
Due to availability of different models, fore- cast produces different plausible scenarios in forecasting since forecasting is a non- deterministic prediction of future events. The scenario with maximum probability value can be considered and it can minimize the penalty due to deviation. Scenario based analysis is an effective tool to predict the future events unlike NWP model based prediction which produces deterministic future events.
A good forecast reduces the uncertainty in prediction to accommodate the variability in wind power generation. Though there is a paucity of historical data and the data related to weather parameters of different wind plants in India, but the del2infinity’s AI based solution with proper choices of models customized for each wind plant can be used for forecasting and scheduling solution which can minimize the penalty due to present regulation and can work even when there is unscheduled variability. As an example, the figure shows the daily- forecast of a 160 MW plant in India, though there are variations in the power generation, the AI based algorithm is useful to forecast the pattern of fluctuations.
For a high penetration of wind in existing grid and to maintain the reliability in the energy distribution, the requirement of forecasting and scheduling in wind power generation at generator level is inevitable and there is need of an accurate forecast and scheduling in wind power for a sustain- able energy future.
Credits: Amit Kumar Das, Director,Del2infinity Energy Consulting