Wind energy conversion technologies, such as advanced wind turbine management and power station scheduling, rely on accurate wind power forecasting to optimize operations. Improving the precision of wind power forecasts is essential to minimize the volatility associated with wind power integration and enhance the competitiveness of wind energy in energy auction markets. Time series forecasting is a method used to estimate future values in a sequence based on historical data. However, traditional wind power forecasting approaches face computational time and complexity limitations. As a result, machine learning researchers have recently focused on overcoming these challenges. We propose an intelligent forecasting model that uses a bidirectional long-short-term memory network (Bi-LSTM) optimized with Bayesian optimization using Tree-structured Parzen Estimators (TPE) to improve the accuracy of the forecasts of short-and long-term wind power. Wind power forecasting (WPF) is classified into two main aspects in the proposed research: multivariate estimation (MVE), which considers input features such as wind direction, speed, temperature, and pressure. In contrast, pure time-series prediction (TSP) analyzes the historical power generation pattern over a specific period for WPF. A sixyear historical wind energy dataset from the NERL lab evaluates the performance of the proposed WPF model. The validation of the proposed method is assessed using root mean squared error (RMSE), mean absolute error (MAE), and R2 score. The findings demonstrate that the proposed TPE-optimized Bi-LSTM model achieves higher MVE and TSP accuracy than the Vanilla LSTM, Stacked LSTM, and Bi-LSTM. Finally, this study gives the findings and prospective recommendations for improving WPF based on a thorough and critical analysis of the results.