Predictions of renewable energy (RE) generation and electricity load are critical to smart grid operation. However, the prediction task remains challenging due to the intermittent and chaotic character of RE sources, and the diverse user behavior and power consumers. This article presents a novel method for the prediction of RE generation and electricity load using improved stacked gated recurrent unit-recurrent neural network (GRU-RNN) for both univariate and multivariate scenarios. First, multiple sensitive monitoring parameters or historical electricity consumption data are selected according to the correlation analysis to form the input data. Second, a stacked GRU-RNN using a simplified GRU is constructed with improved training algorithm based on AdaGrad and adjustable momentum. The modified GRU-RNN structure and improved training method enhance training efficiency and robustness. Third, the stacked GRU-RNN is used to establish an accurate mapping between the selected variables and RE generation or electricity load due to its self-feedback connections and improved training mechanism. The proposed method is verified by using two experiments: prediction of wind power generation using multiple weather parameters and prediction of electricity load with historical energy consumption data. The experimental results demonstrate that the proposed method outperforms state-of-the-art methods of machine learning or deep learning in achieving an accurate energy prediction for effective smart grid operation.

Unsupervised cross-domain fault diagnosis of bearings has practical significance; however, the existing studies still face some problems. For example, transfer diagnosis scenarios are limited to the experimental domain, cross-domain marginal distribution and conditional distribution are difficult to align simultaneously, and each source-domain sample is assigned with equal importance during the domain adaptation process. Aiming at the above-mentioned challenges, this article proposes a novel joint transfer network for unsupervised bearing fault diagnosis from the simulation domain to the experimental domain. The sufficient bearing simulation data containing rich fault label information are used to construct the source domain to reduce the dependence on the resources of laboratory test rigs. An improved loss function embedded with joint maximum mean discrepancy is designed to achieve simultaneous alignments of marginal and conditional distributions across domains in unsupervised scenarios. A weight allocation mechanism for each source-domain sample is developed to suppress negative transfer. Two experimental datasets collected from laboratory test rigs are used as the target domains to validate the effectiveness of the proposed method. The results show that the proposed method is superior to other popular unsupervised cross-domain fault diagnosis methods.