Search

We present a data-driven approach for probabilistic wind power forecasting based on conditional normalizing flow (CNF). In contrast with the existing, this approach is distribution-free (as for non-parametric and quantile-based approaches) and can directly yield continuous probability densities, hence avoiding quantile crossing. It relies on a base distribution and a set of bijective mappings. Both the shape parameters of the base distribution and the bijective mappings are approximated with neural networks. Spline-based conditional normalizing flow is considered owing to its non-affine characteristics. Over the training phase, the model sequentially maps input examples onto samples of base distribution, given the conditional contexts, where parameters are estimated through maximum likelihood. To issue probabilistic forecasts, one eventually maps samples of the base distribution into samples of a desired distribution. Case studies based on open datasets validate the effectiveness of the proposed model, and allows us to discuss its advantages and caveats with respect to the state of the art. [ABSTRACT FROM AUTHOR]

To narrow the width of prediction interval while guaranteeing coverage for probabilistic short term load forecasting, we propose a deep-learning forecasting model based on neural basis expansion analysis (N-BEATS). It takes load data as input, and feed the load sequence into three stacks. Each stack projects the load sequence on a set of basis vectors. Both the basis vectors and the corresponding coefficients are learned by the neural networks. A novel doubly residual stacking strategy is adopted, which decomposes forecasting task into three sub-problems, i.e., pattern characterization tasks corresponding to the stacks, under the assumption that load series can generally be represented by three patterns in subspaces with lower dimensions respectively. It removes redundant information in each stack, which guides the stack to concentrate on learning of one pattern. We further apply conformal quantile regression, which uses the residuals in a held-out validation, to calibrate constructed prediction interval for better theoretical coverage guarantee. Experiments based on load dataset provided by UT Dallas demonstrate improved performance of the proposed model in capturing the characteristics of load and providing narrow prediction intervals with nearly nominal coverage. [ABSTRACT FROM AUTHOR]

Summary: With the development of Energy Internet and the widespread use of smart meters, the application of user‐side data has been attracting more and more attention, one of which is residential short‐term load forecasting (SLTF). The ability of traditional prediction methods is limited due to a large amount of fine‐grained data with high uncertainty, and many researchers have applied deep learning to SLTF thereby. However, perspectives to study and develop deep neural network forecasters are yet to be freed from time domain. Frequency domain analysis provides a sufficient path to assess the periodic characteristics of load. By insights from the spectrum of load data, significant periodic fluctuation characteristics of load are revealed. Based on this, a dual time granularity inputs structured RNNs ensemble STLF model is proposed, which unites two independent input‐featured RNNs to learn more time‐frequency characteristics implied in fine‐grained load. One RNN is mainly responsible for mining time‐domain features of load, the other focuses on learning frequency‐domain features. Case study indicates a positive correlation between capturing more spectrum characteristics and performing a better prediction over accepted criteria, which has been achieved by the proposed model. In addition, experiments show that the proposed forecaster can improve both individual household and aggregated load STLF accuracy. [ABSTRACT FROM AUTHOR]