Hierarchical Structure Based Energy Consumption Forecasting in Top-Down Approach

End-to-End Top-Down Load Forecasting Model for Residential Consumers

“…Many studies have been dedicated to building forecast models for hierarchical load structures [24][25][26]. The hierarchical forecast model follows either a TD or BU approach.…”

“…In BU, forecasts are conducted at lower levels of the hierarchy, which are summed up to obtain an estimate at a higher level. Therefore, the BU approach necessitates employing several forecast models, each corresponding to a sub-level within the hierarchy, and necessitates the need for optimizing each model to achieve the most accurate estimation, whereas in the TD approach, the forecast is performed at the highest level of the hierarchy, which is then distributed down to lower levels using mathematical relationships between different levels of the hierarchy [26]. The TD approach provides better accuracy at the aggregated level but may lead to lower accuracy at lower levels due to the possibility of the loss of information.…”

“…In one of our previous works [26], we tested the efficacy of models based on the TD approach by building a two-stage model to forecast the load consumption of the residential hierarchical structure. It consists of two models, where model A uses features of the aggregated level and is trained to forecast load at the aggregated level.…”

“…Moreover, the input features for the two models are dependent on the individual features of the aggregated level of hierarchy and the level for which the forecast is required, thereby disregarding the relationship between the two levels of the hierarchy i.e., the aggregated level and the cluster for which the forecast is required. This paper builds upon the work presented in [26] and introduces a single model capable of predicting load consumption at all levels of the hierarchy. Furthermore, it incorporates distinctive features that explain the relationship between the aggregated level and the cluster, thereby promoting cross-learning between different levels of the hierarchy.…”

“…The second method calculates the RLP by averaging the daily load profiles of each customer and normalizing it relative to their maximum power value. Both of these methods were evaluated in [26], and it was found that the average method gives better RLPs. This work builds upon those findings, and seasonal RLPs are created for each customer for all four seasons: winter, summer, fall, and spring.…”