• Energy Research

Electricity demand prediction is crucial to ensure the operational safety and cost-efficient operation of the power system. Electricity demand has predominantly been predicted deterministically, while uncertainty analysis has been usually overlooked. To address this research gap, an integrated Neural Facebook Prophet (NFBP) model and Gaussian Kernel Density Estimation (KDE) model is proposed in this paper, as a way to obtain point and interval predictions of electricity demand, quantifying this way the uncertainty in the predictions. First, historical lagged data, created by utilizing the Partial Auto-correlation Function and Mutual Information Test, is applied to train a prediction model based on NFBP, Deep Learning (DL) as well as Statistical Models. Second, the model Prediction Errors (PE) are derived from the difference between actual and predicted values. A splitting strategy based on the mean and standard deviation of PE is proposed. Finally, electricity demand prediction intervals are obtained by applying Gaussian KDE on split PE. To verify the effectiveness of the proposed model, simulation studies are carried out for three prediction horizons on freely available datasets for the Bulimba sub-station in Southeast Queensland, Australia. Compared with DL models (Long-Short Term Memory Network and Deep Neural Network), the Root Mean Square Error of the NFBP model was reduced by 6.1% and 11.3% for 0.5-hr ahead, 22.7% and 26.3% for 6-hr ahead, and 31.8% and 29.9% for daily prediction. In addition, the Prediction Interval normalized Interval width is smaller in magnitude for the proposed NFBP-KDE model compared to other DL and Statistical models

The authors thank data providers, all reviewers and Editor for their thoughtful comments, suggestions and review process. Partial support of this work was through a project PID2020-115454GB-C21 of the Spanish Ministry of Science and Innovation (MICINN) . Prediction of electricity price is crucial for national electricity markets supporting sale prices, bidding strategies, electricity dispatch, control and market volatility management. High volatility, non-stationarity and multi-seasonality of electricity prices make it significantly challenging to estimate its future trend, especially over near real-time forecast horizons. An error compensation strategy that integrates Long ShortTerm Memory (LSTM) network, Convolution Neural Network (CNN) and the Variational Mode Decomposition (VMD) algorithm is proposed to predict the half-hourly step electricity prices. A prediction model incorporating VMD and CLSTM is first used to obtain an initial prediction. To improve its predictive accuracy, a novel error compensation framework, which is built using the VMD and a Random Forest Regression (RF) algorithm, is also used. The proposed VMD-CLSTM-VMD-ERCRF model is evaluated using electricity prices from Queensland, Australia. The results reveal highly accurate predictive performance for all datasets considered, including the winter, autumn, spring, summer, and yearly predictions. As compared with a predictive model without error compensation (i.e., the VMD-CLSTM model), the proposed VMD-CLSTM-VMD-ERCRF model outperforms the benchmark models. For winter, autumn, spring, summer, and yearly predictions, the average Legates and McCabe Index is seen to increase by 15.97%, 16.31%, 20.23%, 10.24%, and 14.03%, respectively, relative to the benchmark models. According to the tests performed on independent datasets, the proposed VMD-CLSTMVMD-ERCRF model can be a practical stratagem useful for short-term, half-hourly electricity price forecasting. Therefore the research outcomes demonstrate that the proposed error compensation framework is an effective decision-support tool for improving the predictive accuracy of electricity price. It could be of practical value to energy companies, energy policymakers and national electricity market operators to develop their insight analysis, electricity distribution and market optimization strategies.

We review the latest modeling techniques and propose new hybrid SAELSTM framework based on Deep Learning (DL) to construct prediction intervals for daily Global Solar Radiation (GSR) using the Manta Ray Foraging Optimization (MRFO) feature selection to select model parameters. Features are employed as potential inputs for Long Short-Term Memory and a seq2seq SAELSTM autoencoder Deep Learning (DL) system in the final GSR prediction. Six solar energy farms in Queensland, Australia are considered to evaluate the method with predictors from Global Climate Models and ground-based observation. Comparisons are carried out among DL models (i.e., Deep Neural Network) and conventional Machine Learning algorithms (i.e., Gradient Boosting Regression, Random Forest Regression, Extremely Randomized Trees, and Adaptive Boosting Regression). The hyperparameters are deduced with grid search, and simulations demonstrate that the DL hybrid SAELSTM model is accurate compared with the other models as well as the persistence methods. The SAELSTM model obtains quality solar energy prediction intervals with high coverage probability and low interval errors. The review and new modelling results utilising an autoencoder deep learning method show that our approach is acceptable to predict solar radiation, and therefore is useful in solar energy monitoring systems to capture the stochastic variations in solar power generation due to cloud cover, aerosols, ozone changes, and other atmospheric attenuation factors.

Wind, being a clean and sustainable resource, boasts environmental advantages. However, its electricity generation faces challenges due to unpredictable variations in wind speed (WS). Accurate predictions of these variations would allow mixed grids to adjust their energy mix in real-time, ensuring overall stability. For this purpose, the paper develops a new hybrid gated additive tree ensemble (H-GATE) model for accurate multi-step-ahead WS predictions. First, the multivariate empirical mode decomposition (MEMD) simultaneously demarcates the multivariate data into intrinsic mode functions (IMFs) and residuals. These components represent underlying trends, periodicity, and stochastic patterns in WS variations. The IMF and residual components are pooled in respective sets, and an opposition-based whale optimization algorithm (OBWOA) is applied for dimensionality reduction. The selected features are used by GATE tuned with Bayesian optimization (BO) to forecast the individual IMF and residual components. The outputs are summed to obtain the final multi-step-ahead WS forecasts. The proposed H-GATE is benchmarked against standalone (S-GATE, S-CLSTM, and S-ABR) and hybrid (H-CLSTM and H-ABR) models. Based on all statistical metrics and diagnostic plots, H-GATE outperforms all comparative models at all forecast horizons, accumulating the lowest mean absolute percentage error (MAPE) of 6.13 - 9.93% (at $t_{L+1}$ ), 8.67 - 14.07% (at $t_{L+2}$ ), and 11.60 - 18.37% (at $t_{L+3}$ ) across all three sites. This novel multi-step-ahead WS forecasting strategy can significantly benefit grid operators by helping anticipate fluctuations in wind power generation. This can assist in optimizing energy dispatch schedules, reducing reliance on backup power sources, and enhancing overall grid stability. Practical implementation of this method can help meet the rising energy demands through renewable wind energy.

Abstract Real-time energy management systems that are designed to support consumer supply and demand spectrums of electrical energy continue to face challenges with respect to designing accurate and reliable real-time forecasts due to the stochasticity of model construction data and the model’s inability to disseminate both the short- and the long-term electrical energy demand (G) predictions. Using real G data from Queensland, Australia’s second largest state, and employing the support vector regression (SVR) model integrated with an improved version of empirical mode decomposition with adaptive noise (ICEEMDAN) tool, this study aims to propose a novel hybrid model: ICEEMDAN-PSO-SVR. Optimization of the model’s weights and biases was performed using the particle swarm optimization (PSO) algorithm. ICEEMDAN was applied to improve the hybrid model’s forecasting accuracy, addressing non-linear and non-stationary issues in time series inputs by decomposing statistically significant historical G data into intrinsic mode functions (IMF) and a residual component. The ICEEMDAN-PSO-SVR model was then individually constructed to forecast IMFs and the residual datasets and the final G forecasts were obtained by aggregating the IMF and residual forecasted series. The performance of the ICEEMDAN-PSO-SVR technique was compared with alternative approaches: ICEEMDAN-multivariate adaptive regression spline (MARS) and ICEEMDAN-M5 model tree, as well as traditional modelling approaches: PSO-SVR, MARS and M5 model tree algorithms. To develop the models, data were partitioned into different subsets: training (70%), validation (15%), and testing (15%), and the tuned forecasting models with near global optimum solutions were applied and evaluated at multiple horizons: short-term (i.e., weekends, working days, whole weeks, and public holidays), and long-term (monthly). Statistical metrics including the root-mean square error (RMSE), mean absolute error (MAE) and their relative to observed means (RRMSE and MAPE), Willmott’s Index (WI), the Legates and McCabe Index ( E LM ) and Nash–Sutcliffe coefficients ( E NS ), were used to assess model accuracy in the independent (testing) period. Empirical results showed that the ICEEMDAN-PSO-SVR model performed well for all forecasting horizons, outperforming the alternative comparison approaches: ICEEMDAN-MARS and ICEEMDAN-M5 model tree and the PSO-SVR, PSO-MARS and PSO-M5 model tree algorithm. Due to its high predictive utility, the two-phase ICEEMDAN-PSO-SVR hybrid model was particularly appropriate for whole week forecasts ( E NS = 0.95 , MAPE = 0.89 % , RRMSE = 1.22 % , and E LM = 0.79 ), and monthly forecasts ( E NS = 0.70 , MAPE = 2.18 % , RRMSE = 3.18 % , and E LM = 0.56 ). The excellent performance of the ICEEMDAN-PSO-SVR hybrid model indicates that the two-phase hybrid model should be explored for potential applications in real-time energy management systems.

Abstract Reliable models that can forecast energy demand (G) are needed to implement affordable and sustainable energy systems that promote energy security. In particular, accurate G models are required to monitor and forecast local electricity demand. However, G forecasting is a multivariate problem, and thus models must employ robust pattern recognition algorithms that can detect subtle variations in G due to causal factors, such as climate variables. Therefore, this study developed an artificial neural network (ANN) model that used climatic variables for 6-hour (h) and daily G forecasting. The input variables included the six most relevant climate variables from Scientific Information for Land Owners (SILO) and 51 Reanalysis variables obtained from the European Centre for Medium-Range Weather Forecast (ECMWF) models. This information was used to forecast G data obtained from the energy utility (Energex) at 8 stations in southeast Queensland, Australia, by utilizing statistically significant lagged cross-correlations of G with its predictor variables. The developed ANN model was then benchmarked against multivariate adaptive regression spline (MARS), multiple linear regression (MLR), and autoregressive integrated moving average (ARIMA) models using various statistical metrics, such as relative root-mean square error (RRMSE%). Additionally, this study developed a hybrid ANN model by combining the forecasts of the ANN, MARS, and MLR models. The bootstrap (B) technique was also used with the hybrid ANN model, creating the B-hybrid ANN, to estimate the forecast uncertainty. According to both forecast horizons, the results indicated that the ANN model was more accurate than the ARIMA, MARS, and MLR models for G forecasting. Furthermore, the hybrid ANN was the most accurate model developed in this research study. For example, at the best site (Redcliffe), the hybrid ANN model generated an RRMSE of 3.85% and 4.37% for the 6-h and daily horizons, respectively. This study found that an ANN model could be used for accurately forecasting G over multiple horizons in southeast Queensland.

Abstract Global advocacy to mitigate climate change impacts on pristine environments, wildlife, ecology, and health has led scientists to design technologies that harness solar energy with remotely sensed, freely available data. This paper presents a study that designed a regionally adaptable and predictively efficient extreme learning machine (ELM) model to forecast long-term incident solar radiation (ISR) over Australia. The relevant satellite-based input data extracted from the Moderate Resolution Imaging Spectroradiometer (i.e., normalized vegetation index, land-surface temperature, cloud top pressure, cloud top temperature, cloud effective emissivity, cloud height, ozone and near infrared-clear water vapour), enriched by geo-temporal input variables (i.e., periodicity, latitude, longitude and elevation) are applied for a total of 41 study sites distributed approximately uniformly and paired with ground-based ISR (target). Of the 41 sites, 26 are incorporated in an ELM algorithm for the design of a universal model, and the remainder are used for model cross-validation. A universally-trained ELM (with training data as a global input matrix) is constructed, and the spatially-deployable model is applied at 15 test sites. The optimal ELM model is attained by trial and error to optimize the hidden layer activation functions for feature extraction and is benchmarked with competitive artificial intelligence algorithms: random forest (RF), M5 Tree, and multivariate adaptive regression spline (MARS). Statistical metrics show that the universally-trained ELM model has very good accuracy and outperforms RF, M5 Tree, and MARS models. With a distinct geographic signature, the ELM model registers a Legates & McCabe's Index of 0.555–0.896 vs. 0.411–0.858 (RF), 0.434–0.811 (M5 Tree), and 0.113–0.868 (MARS). The relative root-mean-square (RMS) error of ELM is low, ranging from approximately 3.715–7.191% vs. 4.907–10.784% (RF), 7.111–11.169% (M5 Tree) and 4.591–18.344% (MARS). Taylor diagrams that illustrate model preciseness in terms of RMS centred difference, error analysis, and boxplots of forecasted vs. observed ISR also confirmed the versatility of the ELM in generating forecasts over heterogeneous, remote spatial sites. This study ascertains that the proposed methodology has practical implications for regional energy modelling, particularly at national scales by utilizing remotely-sensed satellite data, and thus, may be useful for energy feasibility studies at future solar-powered sites. The approach is also important for renewable energy exploration in data-sparse or remote regions with no established measurement infrastructure but with a rich and viable satellite footprint.

The required data was provided by Energex. The study received partial funding from the Ministry of Science and Innovation, Spain (Project ID: PID2020-115454GB-C21). Partial support of this work was through the LATENTIA project PID2022-140786NB-C31 of the Spanish Ministry of Science, Innovation and Universities (MICINNU) . This work presents a Temporal Convolution Network (TCN) model for half-hourly, three-hourly and daily-time step to predict electricity demand ( ) with associated uncertainties for sites in Southeast Queensland Australia. In addition to multi-step predictions, the TCN model is applied for probabilistic predictions of where the aleatoric and epistemic uncertainties are quantified using maximum likelihood and Monte Carlo Dropout methodologies. The benchmarks of TCN model include an attention-based, bi-directional, gated recurrent unit, seq2seq, encoder–decoder, recurrent neural networks and natural gradient boosting models. The testing results show that the proposed TCN model attains the lowest relative root mean square error of 5.336-7.547% compared with significantly larger errors for all benchmark models. In respect to the 95% confidence interval using the Diebold–Mariano test statistic and key performance metrics, the proposed TCN model is better than benchmark models, capturing a lower value of total uncertainty, as well as the aleatoric and epistemic uncertainty. The root mean square error and total uncertainty registered for all of the forecast horizons shows that the benchmark models registered relatively larger errors arising from the epistemic uncertainty in predicted electricity demand. The results obtained for TCN, measured by the quality of prediction intervals representing an interval with upper and lower bound errors, registered a greater reliability factor as this model was likely to produce prediction interval that were higher than benchmark models at all prediction intervals. These results demonstrate the effectiveness of TCN approach in electricity demand modelling, and therefore advocates its usefulness in now-casting and forecasting systems.