A hybrid approach for the techno-economic analysis of wind power plants

This study presents an integrated framework combining design of experiments, the System Advisor Model, and machine learning to improve energy production and economic forecasting for wind power plants. A multi-level full factorial design of experiments generated 2688 scenarios by varying four critical parameters-turbine spacing (3-8 rotor diameters), hub height (80-140 m), rotor diameter (75-110 m), and wind speed (5.66-10.18 m/s)-across eight U.S. regions. These scenarios were simulated in the System Advisor Model to compute five techno-economic outputs: annual energy production, net present value after tax, levelized cost of energy, net capital cost per watt, and total balance-of-system cost. Five machine learning algorithms-neural networks, support vector machines, partial least squares, k-nearest neighbors, and stochastic gradient descent-were then trained on the system advisor model dataset. Among the eight sites, Wyoming yielded the highest energy production (659,434 MWh/year) and the lowest levelized cost of energy (1.59 cent/kWh), while Florida recorded the lowest production (317,876 MWh/year). Neural networks and support vector machines demonstrated superior predictive performance, with neural networks achieving R2 values of 0.964 for the levelized cost of energy and 0.944 for the annual energy production. The sensitivity analysis using a box plot confirmed the robustness and reliability of the developed models. These results demonstrate that integrating the design of experiments, the system advisor model, and machine learning provides accurate, scalable, and robust forecasting tools for the wind power plants' planning and investment decisions.