Heat load forecasting: An online machine learning algorithm for district heating systems

An online machine learning algorithm for heat load forecasting was proposed. These algorithms are increasingly used due to their computational efficiency and their ability to handle changes of the predictive target variable over time. Online bagging creates an ensemble of decision trees. The base model of the ensemble is the Fast Incremental Model Trees with Drift Detection (FIMTDD) Algorithm. When a new instance arrives, FIMT-DD traverses the instance to a terminal node and updates the necessary statistics for this node. Then the algorithm checks if the splitting criterion is satisfied, in order to decide on whether this node should be further expanded. To predict the target value of an instance, FIMT-DD calculates a weighted average of the instance's attributes. The customer heat load in a DH network is calculated by aggregating the heat load of all the nodes in the network. The predictive ability of the proposed algorithm is evaluated by conducting experiments for the two approaches of data aggregation. A key feature of the model is its ability to handle missing values and outliers, which increases the robustness of the model to noise and measurement errors.