Nighttime remote sensing data are considered an important dataset for measuring human social-economic activities. Due to the strong positive correlation between long time series of nighttime remote sensing data and the degree of economic development, it is possible to construct a Gross Domestic Product (GDP) forecasting model using long time series of nighttime remote sensing data to predict GDP. However, traditional statistical methods struggle to accurately and timely reveal the spatial and temporal information of economic parameters and estimate GDP with insufficient precision. How to eliminate data errors in nighttime remote sensing data and construct reliable long time series is currently a major challenge in research. In addition, different GDP forecasting models have different applicable conditions and forecasting accuracy. How to choose the optimal GDP forecasting model is another major challenge in current research. Thus this paper uses nighttime light imagery data from 2016 to 2022, obtained by the Visible Infrared Imaging Radiometer (VIIRS) carried by the National Polar-orbiting Partnership (NPP) satellite as the data source. Errors generated in the data are processed to obtain long time series data suitable for quantitative analysis. Initially, the Random Forest algorithm is employed for predictive research on the dataset. Subsequently, a regression error estimation based on out-of-bag data is proposed. An improved grid search algorithm is used to tune the parameters of the Random Forest model. The grid search algorithm adjusts parameters step by step according to the step size, traverses the parameter range sequentially to find the parameters with the highest accuracy, and uses 5-fold cross-validation for assessment. Additionally, Bayesian optimization is utilized to tune the hyper-parameters in the Random Forest (RF) model. The Bayesian optimization process mainly includes a prior function and an acquisition function. The prior function is implemented using a Gaussian process, employing nested 5F-CV and estimating the model’s generalization ability through an external 5F-CV loop. The internal 5F-CV loop is used to determine the optimal parameters and identify the best parameter model. This optimization method freely combines preset model parameters, conducts a comprehensive search of all parameter combinations, and evaluates the models by analyzing three indicators, comparing the features of different models. The study has established an automatic forecasting system that enables the algorithm model to make accurate predictions automatically based on the input data of the study area. The research results show that after Bayesian optimization and grid optimization, the R2 of the Random Forest algorithm in predicting GDP increased from 0.948 to 0.978 and 0.968, respectively. At the same time, several other evaluation indicators show a significant improvement in forecasting effectiveness. Among them, the Random Forest model improved based on Bayesian optimization has the highest prediction accuracy and the smallest root mean square error value, demonstrating high accuracy and robustness. The research results demonstrate the significant capability of machine learning algorithms and nighttime light index in predicting GDP at the county level. This research framework enriches the application fields and methods of nighttime data and can be used to explore more refined regional development studies in other countries or regions.