Abstract:Transformers have recently gained prominence in long time series forecasting by elevating accuracies in a variety of use cases. Regrettably, in the race for better predictive performance the overhead of model architectures has grown onerous, leading to models with computational demand infeasible for most practical applications. To bridge the gap between high method complexity and realistic computational resources, we introduce the Residual Cyclic Transformer, ReCycle. ReCycle utilizes primary cycle compression to address the computational complexity of the attention mechanism in long time series. By learning residuals from refined smoothing average techniques, ReCycle surpasses state-of-the-art accuracy in a variety of application use cases. The reliable and explainable fallback behavior ensured by simple, yet robust, smoothing average techniques additionally lowers the barrier for user acceptance. At the same time, our approach reduces the run time and energy consumption by more than an order of magnitude, making both training and inference feasible on low-performance, low-power and edge computing devices. Code is available at https://github.com/Helmholtz-AI-Energy/ReCycle
Abstract:This paper presents a new algorithm to extract device profiles fully unsupervised from three phases reactive and active aggregate power measurements. The extracted device profiles are applied for the disaggregation of the aggregate power measurements using particle swarm optimization. Finally, this paper provides a new approach for short term power predictions using the disaggregation data. For this purpose, a state changes forecast for every device is carried out by an artificial neural network and converted into a power prediction afterwards by reconstructing the power regarding the state changes and the device profiles. The forecast horizon is 15 minutes. To demonstrate the developed approaches, three phase reactive and active aggregate power measurements of a multi-tenant commercial building are used. The granularity of data is 1 s. In this work, 52 device profiles are extracted from the aggregate power data. The disaggregation shows a very accurate reconstruction of the measured power with a percentage energy error of approximately 1 %. The developed indirect power prediction method applied to the measured power data outperforms two persistence forecasts and an artificial neural network, which is designed for 24h-day-ahead power predictions working in the power domain.
Abstract:This paper provides a formalization of the energy disaggregation problem for particle swarm optimization and shows the successful application of particle swarm optimization for disaggregation in a multi-tenant commercial building. The developed mathmatical description of the disaggregation problem using a state changes matrix belongs to the group of non-event based methods for energy disaggregation. This work includes the development of an objective function in the power domain and the description of position and velocity of each particle in a high dimensional state space. For the particle swarm optimization, four adaptions have been applied to improve the results of disaggregation, increase the robustness of the optimizer regarding local optima and reduce the computational time. The adaptions are varying movement constants, shaking of particles, framing and an early stopping criterion. In this work we use two unlabelled power datasets with a granularity of 1 s. Therefore, the results are validated in the power domain in which good results regarding multiple error measures like root mean squared error or the percentage energy error can be shown.