Federated Learning is expected to mitigate data privacy risks but introduces more vulnerable surfaces due to its distributed nature. This paper investigates a poisoning attack on federated learning. While recent studies are actively exploring this topic in classification models of learning such as image recognition, there are few studies that address the topic in regression models. This article especially examines the impacts of poisoning attacks on the performance of load forecasting, which has hardly studied yet in academia. To this end, at first, Long Short-Term Memory is implemented on federated learning for load forecasting using publicly available energy data. The first experiment demonstrates how distributed learning affects forecasting performance. Then, we implement two poisoning attacks on the federated learning setting and run experiments to enumerate their impacts on prediction accuracy of load forecasting. Lastly, this paper proposes a spectral clustering algorithm to detect two poisoning attacks and mitigate their impacts and evaluates its performance. Experimental results demonstrate that the proposed algorithm increases forecasting accuracy by 175.9% on a sign flipping attack and by 174.8% on an additive noise attack.