This thesis investigates the feasibility of federated learning (FL) for short-horizon prediction of district-heating substation heat demand from time­series data. Traditional centralized approaches require aggregating multi-dimensional meter readings from all substations into one central database. This consolidation strategy introduces practical constraints concerning data governance requirements, transmission bandwidth limitations, and operational scalability across autonomously managed facilities. Addressing these constraints, this study deploys a distributed learning architecture where individual substations maintain exclusive access to local operational data while contributing parameter updates to a central coordination node for model synthesis. A multilayer perceptron (MLP) serves as the base learner and is trained on hourly sequences derived from substation measurements (kWh and m3/h). The federated approach is implemented using the Flower framework, and two aggregation strategies-Federated Averaging (FedAvg) and adaptive federated optimization (FedAdam)-are evaluated over multiple communication rounds. Performance is assessed using mean absolute error (MAE) on each client's test set, enabling a client-level analysis of convergence and generalization under heterogeneous data distributions. The results indicate that the aggregated global model generally achieves lower MAE than the corresponding client models trained locally in each round, demonstrating the benefit of collaborative learning across substations without pooling raw data. The choice of aggregation strategy affects both convergence and stability: FedAdam provides faster early round improvement and more stable performance across clients in several rounds, whereas FedAvg is more sensitive to client heterogeneity and exhibits larger fluctuations in later rounds. Overall, the study supports FL as a practical approach for distributed substation forecasting and highlights the importance of aggregation strategy and training configuration when deploying FL in real-world district heating settings.