Since the adoption of the Paris Climate Agreement in 2015, there has been a heightened global focus on the issues of global warming and net-zero carbon. This increased interest has prompted companies to shift their management approaches from solely pursuing profits for sustainable growth and competitive advantage to incorporating environmental, social, and governance considerations. Consequently, there is a growing need for energy forecasting networks to alleviate Korea's heavy reliance on energy and facilitate the achievement of net-zero carbon. These networks play a crucial role in predicting uncertain factors such as the volatility of renewable energy availability and electricity prices. To address this need, this paper undertakes two studies aimed at developing an energy prediction network.In the first study, a novel forecasting model for renewable energy (RE) demand was developed, employing both probabilistic and deterministic deep learning algorithms. This approach enabled a comprehensive analysis of energy demand patterns in the RE sector. VAE generated feasible samples, which were subsequently utilized as inputs for BiLSTM to predict RE demand under diverse scenarios. During the data processing phase, present data was transformed into future data through the utilization of a conversion factor, and a meticulous data filtering process was employed to select valid data. To identify the optimal model, various models including Artificial Neural Networks (ANN), Deep Neural Networks (DNN), BiLSTM, LSTM, Gated Recurrent Units (GRU), Autoregressive Integrated Moving Average (ARIMA), Seasonal ARIMA (SARIMA), and Support Vector Regression (SVR) were compared. The evaluation metrics of R2, RMSE, MAE, and MAPE were employed to assess performance, wherein BiLSTM exhibited superior results with values of 0.979, 1,324, 802, and 1.802, respectively. Additionally, the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) were employed to consider both model complexity and performance. AIC and BIC demonstrated optimal performance with values of 28,425 and 41,068, respectively.In the second study, following the identification of the optimal batch model, a dynamically updated prediction model for the Korean System Marginal Price (SMP) was developed utilizing online learning techniques. This approach facilitated real-time adaptation and refinement of the SMP prediction model to enhance its accuracy and reliability. Given that a significant portion of Korea's energy sources are imported, there is a considerable time lag involved in power generation from these sources. To account for this time lag, a time interval between the input feature and the SMP was established and updated every quarter. As the Korean SMP did not exhibit a discernible temporal trend, instead of employing time series models, a comparison was made between simple DNN, DNN, and SVR models, which are multi-input single-output models, to select the best batch model. The selected optimal batch model, the simple DNN, was then employed for online updates. During the update process, the performance of the model was assessed using evaluation metrics, resulting in R2, RMSE, MAE, and MAPE values of 0.924, 7.991, 5.035, and 0.052, respectively.