Short term load forecasting with markovian switching distributed deep belief networks

• A novel distributed DDBN model for short term load forecasting is proposed. The proposed can be trained under a framework, which doesn’t need central controller. Markovian-based switching topology designed to deal with uncertain cyberattacks during neighbour communication. has high potential handling massive and data. In modern power systems, centralised (STLF) methods raise concern on communication requirements reliability when controller undertakes the processing of data solely. As an alternative, avoid problems mentioned above, whilst possible issues accuracy still exist. To address two issues, deep belief networks (DDBN) Markovian accurate STLF, based completely framework. Without governor, dataset separated locally, while obtaining updates through stochastic neighbours consensus procedure, therefore significantly reduced training time. overall network against enhanced by continually topologies. meanwhile, ensure that structure stable such varying topology, gain delicately designed, convergence algorithm theoretically analysed via Lyapunov function. Besides, restricted Boltzmann machines (RBM) unsupervised learning employed initialisation thereby guaranteeing success rate STLF training. GEFCom 2017 competition ISO New England datasets are applied validate effectiveness method. Experiment results demonstrate enhance around 19% better than DBN algorithm.