Adaptive robust polynomial regression for power curve modeling with application to wind power forecasting

Abstract

Wind farm power curve modeling, which characterizes the relationship between meteorological variables and power production, is a crucial procedure for wind power forecasting. In many cases, power curve modeling is more impacted by the limited quality of input data rather than the stochastic nature of the energy conversion process. Such nature may be due the varying wind conditions, aging and state of the turbines, etc. And, an equivalent steady-state power curve, estimated under normal operating conditions with the intention to filter abnormal data, is not sufficient to solve the problem because of the lack of time adaptivity. In this paper, a refined local polynomial regression algorithm is proposed to yield an adaptive robust model of the time-varying scattered power curve for forecasting applications. The time adaptivity of the algorithm is considered with a new data-driven bandwidth selection method, which is a combination of pilot estimation based on blockwise least-squares parabolic fitting and the probability integral transform. The regression model is then extended to a more robust one, in which a new dynamic forgetting factor is defined to make the estimator forget the out-of-date data swiftly and also achieve a better trade-off between robustness against noisy data and time adaptivity. A case study based on a real-world dataset validates the properties of the proposed regression method. Results show that the new method could flexibly respond to abnormal data at different lead times and has better performance than common benchmarks for short-term forecasting.

Wind turbine power curve estimation based on earth mover distance and artificial neural networks

“…Various methods have been widely employed in power curve modeling and can mainly be divided into two categories as parametric and non-parametric methods [6,[11][12][13][14]. The former category mainly contains linearised segmented model [15], polynomial regression [8,16], maximum principle method, dynamic power curve, logistic regression [17] and probabilistic model. On the other hand, the latter one contains copula [18], cubic spline interpolation, neural network [19], fuzzy logic and data mining algorithms [20,21].…”

“…In practice, the theoretical power captured by a wind turbine depends, among other things, on the air density, which is itself a function of temperature, pressure and humidity, which clearly depend on where the turbine is installed [1]. In addition, the power curve is basically nonlinear and non‐stationary because of the fluctuating and stochastic nature of the wind resource, and it also comprises a noise component which represents all the unavailable microscopic interactions [8]. Finally, the power curve is affected by the conditions of the turbine and its associated equipment.…”

On Comparison of Multiple Non-linear Profiles for Process Selection

“…Quality characteristics often appear to be non-linear profile data in industry [4,5]. Polynomial regression models were commonly used to fit non-linear profile data [6,7].…”

A Review of Approaches for the Detection and Treatment of Outliers in Processing Wind Turbine and Wind Farm Measurements

“…The common approach for modelling WTs and WFs is to use power curves (PCs) [7][8][9][10], which typically provide information on the generated power output (P out ) as a function of the input wind speed (WS). These PC-based WT and WF models have numerous applications, e.g., in wind power forecasting [11,12], WT condition monitoring [7,13], WF condition monitoring [14,15], wind energy resource assessment [16][17][18] and general networks analysis [19]. Manufacturers' power curves (Mfr-PCs)…”