the increasing number of distributed energy resources (ders), such as home batteries and electrical vehicles (evs), provides an opportunity for utility companies to develop demand response mechanisms to balance the demand and supply of energy during peak times. however, it is challenging to shave the grid’s peak load efficiently and effectively as it requires accurate energy forecasting and coordinated management of ders. to address this challenge, this thesis proposes a system consisting of an imagebased ensemble prediction model and a multi-agent reinforcement learning (marl) mechanism for demand response (dr) management in smart grids. for the imagebased prediction model, we hypothesize that the approximate curve of the daily power consumption graph has some specific patterns that can be used to separate each day into different groups based on the pattern of the energy consumption curve. to this end, we use a convolution neural network model to classify and extract the features of the curve image. then, we apply the k-means mechanism for image clustering to select better training sets and optimize the forecasting mechanism. our results show an overall improvement in prediction during the season-changing period. the proposed marl mechanism takes the prediction results as input to the agents to coordinate the discharging time of ders to maximize the peak shaving performance. this mechanism requires centralized training and allows distributed execution. the system’s evaluations and experiments are conducted on a real-life dataset, and our results show the proposed system’s effectiveness.