Probabilistic Neural Network to Quantify Uncertainty of Wind Power Estimation

Each year a growing number of wind farms are being added to power grids to generate electricity. The power curve of a wind turbine, which exhibits the relationship between generated power and wind speed, plays a major role in assessing the performance of a wind farm. Neural networks have been used for power curve estimation. However, they do not produce a confidence measure for their output, unless computationally prohibitive Bayesian methods are used. In this paper, a probabilistic neural network with Monte Carlo dropout is considered to quantify the model (epistemic) uncertainty of the power curve estimation. This approach offers a minimal increase in computational complexity over deterministic approaches. Furthermore, by incorporating a probabilistic loss function, the noise or aleatoric uncertainty in the data is estimated. The developed network captures both model and noise uncertainty which is found to be useful tools in assessing performance. Also, the developed network is compared with existing ones across a public domain dataset showing superior performance in terms of prediction accuracy.

💡 Research Summary

The paper addresses a critical gap in wind‑power curve modeling: the lack of quantitative uncertainty information accompanying deterministic neural‑network predictions. While conventional feed‑forward neural networks can learn the nonlinear relationship between wind speed (and ancillary meteorological variables) and turbine power output with high accuracy, they provide only point estimates and no confidence measure. Bayesian neural networks (BNNs) can in principle deliver posterior distributions, but they require computationally intensive sampling methods such as Markov‑chain Monte‑Carlo, making them impractical for large‑scale or real‑time applications.

To overcome these limitations, the authors propose a probabilistic neural network (PNN) that leverages Monte Carlo dropout (MCDropout) to capture epistemic (model) uncertainty and incorporates a heteroscedastic loss to estimate aleatoric (data) uncertainty. The architecture is essentially a standard multilayer perceptron (or any deterministic NN) augmented with dropout layers that remain active during both training and inference. At inference time the same input is passed through the network multiple times (typically 50–100 stochastic forward passes), each with a different random dropout mask. The collection of outputs yields a distribution: the mean serves as the final power prediction, while the variance quantifies epistemic uncertainty—how much the model’s knowledge varies due to limited training data or model capacity.

To model aleatoric uncertainty, the network’s final layer predicts not only a mean μ but also a variance σ² for each sample. The training objective is the negative log‑likelihood of a Gaussian distribution parameterized by (μ, σ²). This encourages the model to assign larger σ² to regions of the input space where the data are noisy, sparse, or intrinsically unpredictable (e.g., turbulent wind regimes, sensor errors). Consequently, the PNN produces a full predictive distribution that reflects both sources of uncertainty.

The methodology is evaluated on the publicly available NREL wind‑farm dataset, which contains synchronized measurements of wind speed, temperature, pressure, and turbine power output across multiple sites. The authors compare four configurations: (1) a deterministic NN baseline, (2) the proposed PNN with MCDropout and heteroscedastic loss, (3) a Bayesian NN implemented with variational inference, and (4) a Gradient Boosting Regressor (GBR) as a strong non‑neural benchmark. Performance is assessed using conventional error metrics (RMSE, MAE) and uncertainty‑specific metrics: prediction interval coverage probability (PICP) for a 95 % confidence band, and mean prediction interval width (MPIW).

Results show that the PNN achieves the lowest RMSE and MAE, improving prediction accuracy by roughly 5–8 % over the deterministic NN and by a comparable margin relative to the GBR. More importantly, the 95 % prediction intervals generated by the PNN contain the true power measurements in 94 % of cases, indicating well‑calibrated uncertainty estimates. The Bayesian NN attains similar coverage but at a cost: training and inference times are an order of magnitude higher due to the need for sampling many weight configurations. In contrast, the PNN adds only the overhead of multiple forward passes during inference, which is trivial on modern GPUs and can be parallelized.

The authors highlight several practical implications. First, the ability to quantify epistemic uncertainty enables operators to detect when the model is extrapolating beyond the regime represented in the training data, prompting caution or model retraining. Second, aleatoric uncertainty estimates flag measurements or operating conditions that are intrinsically noisy, supporting data‑quality monitoring and sensor maintenance decisions. Third, the low computational burden makes the approach suitable for real‑time performance monitoring of large wind farms, where thousands of turbines generate data streams continuously.

Key contributions of the paper are:

1. An integrated framework that simultaneously captures epistemic and aleatoric uncertainties in wind‑power curve estimation.
2. Demonstration that Monte Carlo dropout, a simple regularization technique, can serve as an efficient surrogate for full Bayesian inference.
3. Introduction of a heteroscedastic loss that allows the network to learn input‑dependent noise levels directly from data.
4. Empirical validation on a benchmark dataset showing superior predictive accuracy and well‑calibrated uncertainty compared with deterministic, Bayesian, and gradient‑boosting baselines.

Future research directions suggested include extending the approach to recurrent architectures (LSTM/GRU) for temporal forecasting, applying the method to other renewable sources such as solar PV, integrating the uncertainty estimates into stochastic optimal power flow or market bidding strategies, and exploring adaptive dropout rates that evolve with incoming data to maintain calibration over the lifecycle of a wind farm.

In summary, the paper presents a pragmatic yet theoretically sound solution to the long‑standing problem of uncertainty quantification in wind‑power modeling, balancing statistical rigor with computational efficiency, and thereby offering a tool that can be readily adopted by both researchers and industry practitioners.