A data-driven deep sequence-to-sequence long-short memory method along with a gated recurrent neural network for wind power forecasting

Abstract Large amounts of wind power generation have an impact not only on energy markets but also on wholesale and retail market designs. Simultaneously, technological issues arise as a result of the need to ensure the smooth operation of the power grid. In the long run, high-quality wind data series are required to generate model results that lead to robust policy advice. Several techniques are commonly used to forecast short-term, nonstationary, and nonlinear wind speeds. These techniques are lacking in model optimization and data processing abilities. This lack of expertise is posing a significant challenge for reliable and stable wind power forecasting. Furthermore, the medium and long-term forecasting criteria for models (e.g., robustness, accuracy, speed) are higher, making it difficult to obtain reliable results. This study used the extended deep sequence-to-sequence long short-term memory regression (STSR-LSTM) model for time-series wind power forecasting to overcome these challenges for accurate forecasting decisions. The statistical-learning technique is used to improve the dependability of the derived features as well as the expected performance. Three different locations/sites were used and analyzed (e.g., Belgian, DSO-Connected, Elia), three different forecasting classes (e.g., week-ahead forecast, day-ahead forecast, and most recent forecast), three different seasons (e.g., monthly, seasonal, and annual), and three experimental setups. The performance of Deep STSR-LSTM is validated using two existing models and four performance evaluation indexes. Despite the fact that the input wind power load curve had a lot of variation, we were able to achieve higher forecast accuracy. Using different sites and classes, similar forecast accuracy was achieved for each season (e.g., monthly, seasonal, and annual).