Search

Anomalous data can negatively impact energy forecasting by causing model parameters to be incorrectly estimated. This paper presents two approaches for the detection and imputation of anomalies in time series data. Autoregressive with exogenous inputs (ARX) and artificial neural network (ANN) models are used to extract the characteristics of time series. Anomalies are detected by performing hypothesis testing on the extrema of the residuals, and the anomalous data points are imputed using the ARX and ANN models. Because the anomalies affect the model coefficients, the data cleaning process is performed iteratively. The models are re-learned on “cleaner” data after an anomaly is imputed. The anomalous data are reimputed to each iteration using the updated ARX and ANN models. The ARX and ANN data cleaning models are evaluated on natural gas time series data. This paper demonstrates that the proposed approaches are able to identify and impute anomalous data points. Forecasting models learned on the unclean data and the cleaned data are tested on an uncleaned out-of-sample dataset. The forecasting model learned on the cleaned data outperforms the model learned on the unclean data with 1.67% improvement in the mean absolute percentage errors and a 32.8% improvement in the root mean squared error. Existing challenges include correctly identifying specific types of anomalies such as negative flows. [ABSTRACT FROM PUBLISHER]

Very short-term load forecasting predicts the loads 1 h into the future in 5-min steps in a moving window manner based on real-time data collected. Effective forecasting is important in area generation control and resource dispatch. It is however difficult in view of the noisy data collection process and complicated load features. This paper presents a method of wavelet neural networks with data pre-filtering. The key idea is to use a spike filtering technique to detect spikes in load data and correct them. Wavelet decomposition is then used to decompose the filtered loads into multiple components at different frequencies, separate neural networks are applied to capture the features of individual components, and results of neural networks are then combined to form the final forecasts. To perform moving forecasts, 12 dedicated wavelet neural networks are used based on test results. Numerical testing demonstrates the effects of data pre-filtering and the accuracy of wavelet neural networks based on a data set from ISO New England. [ABSTRACT FROM PUBLISHER]

Short-run forecasting of electricity prices has become necessary for power generation unit schedule, since it is the basis of every profit maximization strategy. In this article a new and very easy method to compute accurate forecasts for electricity prices using mixed models is proposed. The main idea is to develop an efficient tool for one-step-ahead forecasting in the future, combining several prediction methods for which forecasting performance has been checked and compared for a span of several years. Also as a novelty, the 24 hourly time series has been modelled separately, instead of the complete time series of the prices. This allows one to take advantage of the homogeneity of these 24 time series. The purpose of this paper is to select the model that leads to smaller prediction errors and to obtain the appropriate length of time to use for forecasting. These results have been obtained by means of a computational experiment. A mixed model which combines the advantages of the two new models discussed is proposed. Some numerical results for the Spanish market are shown, but this new methodology can be applied to other electricity markets as well. [ABSTRACT FROM AUTHOR]

Artificial neural networks (ANNs) have been widely applied in electricity price forecasts due to their nonlinear modeling capabilities. However, it is well known that in general, traditional training methods for ANNs such as back-propagation (BP) approach are normally slow and it could be trapped into local optima. In this paper, a fast electricity market price forecast method is proposed based on a recently emerged learning method for single hidden layer feed-forward neural networks, the extreme learning machine (ELM), to overcome these drawbacks. The new approach also has improved price intervals forecast accuracy by incorporating bootstrapping method for uncertainty estimations. Case studies based on chaos time series and Australian National Electricity Market price series show that the proposed method can effectively capture the nonlinearity from the highly volatile price data series with much less computation time compared with other methods. The results show the great potential of this proposed approach for online accurate price forecasting for the spot market prices analysis. [ABSTRACT FROM PUBLISHER]

Despite its great importance, there has been no general consensus on how to model the trends in time-series data. Compared to traditional approaches, neural networks (NNs) have shown some promise in time-series forecasting. This paper investigates how to best model trend time series using NNs. Four different strategies (raw data, raw data with time index, detrending, and dif- ferencing) are used to model various trend patterns (linear, non- linear, deterministic, stochastic, and breaking trend). We find that with NNs differencing often gives meritorious results regardless of the underlying data generating processes (DGPs). This finding is also confirmed by the real gross national product (GNP) series. [ABSTRACT FROM AUTHOR]

The article presents and summarizes the research results of the UTC-UTC(PL) differences predicting procedure for the Polish Time Scale UTC(PL) developed at the Institute of Metrology, Electronics and Computer Science of the University of Zielona Góra. The research has been conducted in close cooperation with the Time and Frequency Laboratory of Central Office of Measures, responsible for the realization of the UTC(PL). The planned experiment covered a 50-month period, in which the UTC(PL) timescale has been carried out using, as a master clock, atomic clocks with different properties: commercial cesium atomic clock 5071A Opt. 001 (first case) and VCH-1003M hydrogen maser (second and third cases). In case of the developed predicting procedure by means of group method of data handling (GMDH)-type neural network (NN), the transitional period (second case) in which the master clock had been replaced by the new clock, is also very important. For GMDH-type NN, input data as a TS1 and TS2 time series have been prepared, based on the already-known differences of UTC(PL) from UTC and UTC rapid timescales and on determined differences between UTC(PL) and the master clock steering the UTC(PL). The method of preparing TS1 and TS2 time series is an important part of the developed procedure for predicting UTC(PL) differences from UTC, before they will be calculated and published by fr. Bureau International des Poids et Mesures (BIPM). The obtained research results confirmed that the developed predicting procedure by means of GMDH-type NN ensures a proper quality of maintaining the national timescale. The best quality of UTC(PL) predicting has been achieved for the third case, both for data prepared according to the TS1 and TS2 time series. This is confirmed by the calculated values of the prediction quality measures and the Modified Allan deviation. On the other hand, the second case confirmed the possibility of using the proprietary development of the TS2 time series to predict UTC-UTC(k) differences. This time series is a very good alternative for the TS1 time series, when the amount of data in this series is insufficient due to replacement of the clock or its failure. [ABSTRACT FROM AUTHOR]

This letter proposes a novel granule computing-based framework for prediction intervals (PIs) construction of solar irradiance time series that has significant impacts on solar power production. Distinguished from most existing methods, the new framework can address both stochastic and knowledge uncertainties in constructing PIs. The proposed method has proved to be highly effective in terms of both reliability and sharpness through a real case study using measurement data obtained from Hong Kong Observatory. [ABSTRACT FROM PUBLISHER]

Rapidly evolving businesses generate massive amounts of time-stamped data sequences and cause a demand for both univariate and multivariate time series forecasting. For such data, traditional predictive models based on autoregression are often not sufficient to capture complex nonlinear relationships between multidimensional features and the time series outputs. In order to exploit these relationships for improved time series forecasting while also better dealing with a wider variety of prediction scenarios, a forecasting system requires a flexible and generic architecture to accommodate and tune various individual predictors as well as combination methods. In reply to this challenge, an architecture for combined, multilevel time series prediction is proposed, which is suitable for many different universal regressors and combination methods. The key strength of this architecture is its ability to build a diversified ensemble of individual predictors that form an input to a multilevel selection and fusion process before the final optimized output is obtained. Excellent generalization ability is achieved due to the highly boosted complementarity of individual models further enforced through cross-validation-linked training on exclusive data subsets and ensemble output postprocessing. In a sample configuration with basic neural network predictors and a mean combiner, the proposed system has been evaluated in different scenarios and showed a clear prediction performance gain. [ABSTRACT FROM AUTHOR]