<p>Forecasts of renewable energy generation provide information for maintaining grid stability, for planning short- and long- term investments in renewable energy, and for optimizing renewable energy pricing. This paper compares three time series models on the accuracies of forecasting the annual U.S. renewable energy generations where the historical sample data sizes are small. The models include two probability distribution-based models, the structural time series (STS) model and the ARIMA model, and a distribution-free model, the Grey system model. Using the annual U.S. energy generation data in Terawatt-hours from Hydroelectric (1965–2023), ‘Others’ (the combined total of geothermal energy and bioenergy, 1965–2023), Solar (1989–2023), and Wind (1989–2023), this paper compares forecasting accuracy of these models based on MSE, MAE and MAPE. The results show that a Grey system model achieves the highest forecasting accuracy for Hydroelectric data and a STS model achieves the highest accuracy for ‘Others’, Solar, and Wind data. And then, the annual U.S. renewable energy generations from year 2024 to 2027 are forecasted using the best performed model for each renewable energy source. The results indicate that both Solar and Wind generation are expected to be sharply increasing, while Hydroelectric is expected to be almost standstill and ‘Others’ is expected to be mildly decreasing. Finally for each energy source, this paper identifies a model which best fits the data based on the Akaike Information Criterion (AIC) among valid STS and ARIMA models.</p>