TL;DR
This paper introduces a novel framework for assessing the reliability of power systems with significant wind integration. By using neural networks trained with a multi-objective genetic algorithm, the authors provide not just point estimates but also interval predictions for wind power and load, leading to more robust and informative estimates of energy not supplied (EENS). This approach better accounts for inherent uncertainties, offering a more realistic picture of system adequacy.
What Problem Does This Paper Address?
Assessing the adequacy of power systems, particularly those with a high penetration of variable renewable energy sources like wind, is a complex challenge. Traditional methods often rely on point estimates for crucial inputs like wind power generation and electricity demand. However, these point estimates fail to capture the inherent uncertainties associated with weather patterns and load fluctuations, leading to potentially oversimplified or misleading assessments of system reliability. This can result in inadequate planning and operational decisions. The paper aims to bridge this gap by developing a framework that explicitly quantifies and incorporates these uncertainties into the adequacy assessment process, specifically focusing on the Expected Energy Not Supplied (EENS) index.
Methods
The core of the paper's methodology lies in its approach to predicting wind power and load, and then using these predictions to estimate EENS.
- Prediction Intervals (PIs) for Wind Power and Load: The authors employ multi-perceptron artificial neural networks (NNs) to forecast both point values and prediction intervals (PIs) for 1-hour ahead wind power and load. To optimize the NN training, they utilize the Non-Dominated Sorting Genetic Algorithm-II (NSGA-II). This multi-objective genetic algorithm is used to simultaneously maximize the PI coverage probability (PICP) and minimize the PI width (PIW), aiming for PIs that are both accurate and narrow.
- Interval-Based EENS Estimation: The paper proposes two strategies for estimating EENS using the generated PIs for load and wind power:
- Interval-based Interval Estimation: This method directly applies interval arithmetic to the EENS formula. It treats the predicted load and wind power as intervals and performs all calculations accordingly, resulting in an interval estimate for EENS. This approach directly propagates the uncertainty from the input predictions to the output EENS.
- Interval-based Point Estimation: This strategy uses the PIs to derive a single point estimate for EENS. It leverages the probability density function of uniform random variables within the predicted intervals to calculate an expected value of EENS. This method aims to provide a more interpretable single value while still accounting for input uncertainties.
- Monte Carlo Simulation: For the adequacy assessment, a sequential Monte Carlo simulation is used to model the failure and repair behavior of conventional generation units, which are represented by two-state models.
Datasets
- Wind Speed Data: Hourly wind speed data for the Regina, Saskatchewan, Canada region, spanning a 9-year period (January 1, 2003, to December 31, 2011), were used. This data was transformed into hourly mean wind speed values for a one-year time series.
- Load Data: Hourly load data was generated using the chronological annual load curve from the IEEE Reliability Test System (RTS) [10], scaled to a specific annual peak load value.
- RBTS System Data: The Roy Billinton Test System (RBTS) was used as the case study. This system consists of conventional generation units, and a wind farm with 20 identical WTG units was added. Reliability data for the conventional units was sourced from [25].
Key Insights and Ideas
- Uncertainty Quantification is Crucial: The paper strongly emphasizes that simply using point estimates for wind power and load can lead to an incomplete understanding of system adequacy. Quantifying and incorporating these uncertainties is essential for robust decision-making.
- Interval Predictions Offer Richer Information: Providing interval estimates for EENS, rather than just point estimates, offers a more comprehensive view. It highlights the range of possible outcomes and the impact of input uncertainties, allowing decision-makers to understand the "worst-case" and "best-case" scenarios.
- NNs and NSGA-II are Powerful Tools: The combination of neural networks for prediction and NSGA-II for optimizing PI quality demonstrates an effective approach to handling complex, non-linear relationships and multi-objective optimization in power system reliability.
- Interval-Based Point Estimation as a Compromise: The interval-based point estimation method offers a practical way to present a single, interpretable EENS value that still benefits from the uncertainty quantification provided by PIs.
Key Results
- The proposed framework successfully generated interval predictions for wind power and load, with the NSGA-II optimization achieving a good balance between PI coverage probability and width.
- The interval-based EENS estimation methods (both interval and point) provided results that were more informative and better reflected the actual EENS compared to traditional point-estimate-based methods.
- The "EENS point" (interval-based point estimate) was found to be the closest to the "EENS actual" (calculated using actual data, not feasible in real-time) on average, indicating its reliability and accuracy in capturing system adequacy under uncertainty.
- The study showed that different scenarios of load and wind levels, as well as component failures, significantly impact EENS, and the proposed methods can effectively capture these variations.
Limitations and Future Work
- Subjective Selection of Pareto Solutions: The paper acknowledges that selecting a specific solution from the Pareto front for NN training is often done subjectively. Future work could explore more objective criteria for this selection.
- Uniform Probability Assumption: The interval-based estimation methods rely on the assumption of uniform probability distribution for load and generation within their predicted intervals. While stated as a natural approach in the absence of further information, this assumption might not always hold true in real-world scenarios.
- Computational Cost: While not explicitly detailed as a limitation, the use of NSGA-II and Monte Carlo simulations suggests that the computational cost of this approach might be higher than traditional methods.
- Future Work Suggestions: The authors suggest that the presented expected annual energy not supplied can be integrated with cost models to aid operational and strategic planning decisions. They also highlight the potential for using these results to identify components contributing most to unsupplied energy and to inform decisions about adding new generation units.
Summary updated