Multi-step Forecasting Research Articles

Explainable machine learning model for multi-step forecasting of reservoir inflow with uncertainty quantification

We propose an explainable machine learning (ML) model with uncertainty quantification (UQ) to improve multi-step reservoir inflow forecasting. Traditional ML methods have challenges in forecasting inflows multiple days ahead, and lack explainability and UQ. To address these limitations, we introduce an encoder–decoder long short-term memory (ED-LSTM) network for multi-step forecasting, employ the SHapley Additive exPlanation (SHAP) technique for understanding the influence of hydrometeorological factors on inflow prediction, and develop a novel UQ method for prediction trustworthiness. We apply these methods to forecast 7-day inflow in snow-dominant and rain-driven reservoirs. The results demonstrate the effectiveness of the ED-LSTM model, with high forecasting accuracy for short lead times. Our UQ method provides reliable uncertainty estimates, covering 90% of data with a 90% confidence level. The SHAP analysis reveals the importance of historical inflow and precipitation as influential factors. These findings and methods may support reservoir operators in optimizing water resources management decisions.

Robust multi-step wind speed forecasting based on a graph-based data reconstruction deep learning method

Against global warming, wind energy has increasingly become a stable form of power supply. Accurate prediction of wind speed is crucial for turbine control and wind farm dispatch, contributing to stable and continued wind energy utilization. However, it is very difficult to accomplish satisfactory wind speed forecasting, especially multi-step forecasting, due to the stochastic, random and volatile characteristic of wind speed series with complex fluctuations. This study is elaborated to propose a novel method based on “data graph” reconstruction and long short term memory (LSTM) network, to achieve accurate and robust multi-step wind speed forecasting. To obtain implicit correlations between wind speed series, the time series data are reconstructed into matrices like “data graphs”. To achieve better computing efficiency and forecasting accuracy, the convolutional neural network (CNN) is adopted to extract the features in the “data graphs”. Then the extracted data graph features are imported into the bidirectional-LSTM (bi-LSTM) network module, which takes the input in forward and backward directions, respectively, to extract more temporal information. For adequate performance assessment, experiments are carried out on data sets of an actual wind farm located in a mountainous region in China. The results show that the proposed method outperforms classic meta-model based and hybrid-model based methods in accuracy, stability and computing efficiency. The results reveal that the data graph reconstruction combined with the CNN is able to extract the hidden features of the wind speed time series data. The proposed CNN-bi-LSTM method effectively improve the multi-step wind speed prediction accuracy.

Variation Trend Prediction of Dam Displacement in the Short-Term Using a Hybrid Model Based on Clustering Methods

The deformation behavior of a dam can comprehensively reflect its structural state. By comparing the actual response with model predictions, dam deformation prediction models can detect anomalies for effective advance warning. Most existing dam deformation prediction models are implemented within a single-step prediction framework; the single-time-step output of these models cannot represent the variation trend in the dam deformation, which may contain important information on dam evolution during the prediction period. Compared with the single value prediction, predicting the tendency of dam deformation in the short term can better interpret the dam’s structural health status. Aiming to capture the short-term variation trends of dam deformation, a multi-step displacement prediction model of concrete dams is proposed by combining the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm, the k-harmonic means (KHM) algorithm, and the error minimized extreme learning machine (EM-ELM) algorithm. The model can be divided into three stages: (1) The CEEMDAN algorithm is adopted to decompose dam displacement series into different signals according to their timing characteristics. Moreover, the sample entropy (SE) method is used to remove the noise contained in the decomposed signals. (2) The KHM clustering algorithm is employed to cluster the denoised data with similar characteristics. Furthermore, the sparrow search algorithm (SSA) is utilized to optimize the KHM algorithm to avoid the local optimal problem. (3) A multi-step prediction model to capture the short-term variation of dam displacement is established based on the clustered data. Engineering examples show that the model has good prediction performance and strong robustness, demonstrating the feasibility of applying the proposed model to the multi-step forecasting of dam displacement.

Forecasting crude oil futures using an ensemble model including investor sentiment and attention

PurposeSince crude oil is crucial to the nation's economic growth, crude oil futures are closely related to many other markets. Accurate forecasting can offer investors trustworthy guidance. Numerous studies have begun to consider creating new metrics from social networks to improve forecasting models in light of their rapid development. To improve the forecasting of crude oil futures, the authors suggest an integrated model that combines investor sentiment and attention.Design/methodology/approachThis study first creates investor attention variables using Baidu search indices and investor sentiment variables for medium sulfur crude oil (SC) futures by collecting comments from financial forums. The authors feed the price series into the NeuralProphet model to generate a new feature set using the output subsequences and predicted values. Next, the authors use the CatBoost model to extract additional features from the new feature set and perform multi-step predictions. Finally, the authors explain the model using Shapley additive explanations (SHAP) values and examine the direction and magnitude of each variable's influence.FindingsThe authors conduct forecasting experiments for SC futures one, two and three days in advance to evaluate the effectiveness of the proposed model. The empirical results show that the model is a reliable and effective tool for predicting, and including investor sentiment and attention variables in the model enhances its predictive power.Research limitations/implicationsThe data analyzed in this paper span from 2018 through 2022, and the forecast objectives only apply to futures prices for those years. If the authors alter the sample data, the experimental process must be repeated, and the outcomes will differ. Additionally, because crude oil has financial characteristics, its price is influenced by various external circumstances, including global epidemics and adjustments in political and economic policies. Future studies could consider these factors in models to forecast crude oil futures price volatility.Practical implicationsIn conclusion, the proposed integrated model provides effective multistep forecasts for SC futures, and the findings will offer crucial practical guidance for policymakers and investors. This study also considers other relevant markets, such as stocks and exchange rates, to increase the forecast precision of the model. Furthermore, the model proposed in this paper, which combines investor factors, confirms the predictive ability of investor sentiment. Regulators can utilize these findings to improve their ability to predict market risks based on changes in investor sentiment. Future research can improve predictive effectiveness by considering the inclusion of macro events and further model optimization. Additionally, this model can be adapted to forecast other financial markets, such as stock markets and other futures products.Originality/valueThe authors propose a novel integrated model that considers investor factors to enhance the accuracy of crude oil futures forecasting. This method can also be applied to other financial markets to improve their forecasting efficiency.

Evaluation of models for multi-step forecasting of hand, foot and mouth disease using multi-input multi-output: A case study of Chengdu, China.

Hand, foot and mouth disease (HFMD) is a public health concern that threatens the health of children. Accurately forecasting of HFMD cases multiple days ahead and early detection of peaks in the number of cases followed by timely response are essential for HFMD prevention and control. However, many studies mainly predict future one-day incidence, which reduces the flexibility of prevention and control. We collected the daily number of HFMD cases among children aged 0-14 years in Chengdu from 2011 to 2017, as well as meteorological and air pollutant data for the same period. The LSTM, Seq2Seq, Seq2Seq-Luong and Seq2Seq-Shih models were used to perform multi-step prediction of HFMD through multi-input multi-output. We evaluated the models in terms of overall prediction performance, the time delay and intensity of detection peaks. From 2011 to 2017, HFMD in Chengdu showed seasonal trends that were consistent with temperature, air pressure, rainfall, relative humidity, and PM10. The Seq2Seq-Shih model achieved the best performance, with RMSE, sMAPE and PCC values of 13.943~22.192, 17.880~27.937, and 0.887~0.705 for the 2-day to 15-day predictions, respectively. Meanwhile, the Seq2Seq-Shih model is able to detect peaks in the next 15 days with a smaller time delay. The deep learning Seq2Seq-Shih model achieves the best performance in overall and peak prediction, and is applicable to HFMD multi-step prediction based on environmental factors.

Dynamic adaptive encoder-decoder deep learning networks for multivariate time series forecasting of building energy consumption

Accurate energy consumption prediction models can bring tremendous benefits to building energy efficiency, where the use of data-driven models allows models to be trained based on historical data and to obtain better prediction performance. However, a severe problem has been overlooked where the dynamic spatiotemporal dependence among buildings is often not considered. To settle this issue, this study proposes an asymmetric encoder-decoder learning framework where the spatial relationships and time-series features between multiple buildings are extracted by a convolutional neural network and a gated recurrent neural network to form new input data in the encoder. The decoder then makes predictions based on the input data with an attention mechanism. The model is trained and tested with energy consumption data from 10 different buildings at a university in China. The results show that our proposed model maintains high accuracy using small datasets while predicting different types of buildings, achieving an average R2 of 0.964. Additionally, the accuracy loss in multi-step forecasting is considerably ameliorated by maintaining an average R2 of 0.915 predicting 3 steps ahead of time. Benefiting from the asymmetric encoder-decoder structure, the proposed algorithm resolves the problem of accuracy loss using deep learning in a small dataset and multi-step time series forecasting. Further assisted by the anomaly detection function, the algorithm serves as a reliable guide for the user to effectively manage and control energy consumption in buildings.

Flexible renewable energy planning based on multi-step forecasting of interregional electricity supply and demand: Graph-enhanced AI approach

Jeju Island faces curtailment problems when renewable energy generation exceeds the facility capacity. In addition, the renewable energy system cannot achieve consistent operation because of the uncertainty of renewable energy sources (RES) and their complex spatiotemporal patterns. Climate change and extreme meteorological phenomena also affect the energy system of the island. Therefore, it is essential to maximize RES exploitation to accomplish carbon-free energy production. This study developed a rolling horizon optimization–based flexible renewable energy distribution planning (FxRE-Plan) based on an artificial intelligence-driven spatiotemporal RES prediction model. An adaptive graph convolutional recurrent network (AGCRN) was used to develop a spatiotemporal prediction model. Then, FxRE-Plan was determined using a multiobjective genetic algorithm to maximize the RES potential under climate change scenarios. The FxRE-Plan was evaluated and compared with the current operation of the electric grid in Jeju Island through economic, environmental cost, and climate change vulnerability assessment index. The results of this research show that the AGCRN–based spatiotemporal prediction model exhibited superior prediction performance, achieving a 95% R2. In addition, the FxRE-Plan reduced greenhouse gas emissions by approximately 11.5% at maximum and climate change vulnerability. Hence, the proposed FxRE-Plan can be used to solve curtailments in real electric grids by pursuing carbon-free islands.

El Niño Index Prediction Based on Deep Learning with STL Decomposition

ENSO is an important climate phenomenon that often causes widespread climate anomalies and triggers various meteorological disasters. Accurately predicting the ENSO variation trend is of great significance for global ecosystems and socio-economic aspects. In scientific practice, researchers predominantly employ associated indices, such as Niño 3.4, to quantitatively characterize the onset, intensity, duration, and type of ENSO events. In this study, we propose the STL-TCN model, which combines seasonal-trend decomposition using locally weighted scatterplot smoothing (LOESS) (STL) and temporal convolutional networks (TCN). This method uses STL to decompose the original time series into trend, seasonal, and residual components. Each subsequence is then individually predicted by different TCN models for multi-step forecasting, and the predictions from all models are combined to obtain the final result. During the verification period from 1992 to 2022, the STL-TCN model effectively captures index features and improves the accuracy of multi-step forecasting. In historical event simulation experiments, the model demonstrates advantages in capturing the trend and peak intensity of ENSO events.

Forecasting Regional Energy Consumption via Jellyfish Search-Optimized Convolutional-Based Deep Learning

The energy sector must achieve a delicate balance between energy supply and demand. Highly accurate energy consumption forecasts can help plant operators achieve this goal. In this study, various techniques from three artificial intelligence categories, namely, convolutional neural networks (CNNs), machine learning (ML), and time-series deep learning (DL), were applied to predict short-term regional energy consumption from a power company. An image conversion process was proposed to utilize the powerful image-processing capabilities of CNNs, in which numerical matrices of input values created by the sliding-window technique are encoded into two-dimensional grayscale images as input for CNN models. The proposed method outperformed conventional numerical input methods. Then, the best ML and time-series DL models were combined into ensemble models with ImageNet-winning CNN architectures based on the rationale of utilizing the strengths of each approach and compensating for their limitations. However, according to the numerical experiments for the regional energy consumption prediction, the ensemble models had lower performance than CNNs alone. Various leading models performing single-step or multistep forecasting and models that use seasonal or annual data were constructed. The performance of all the constructed models was evaluated, and the most efficient models were determined as proposed models. Finally, a nature-inspired optimization algorithm, Jellyfish Search (JS), was used to fine-tune the proposed models’ hyperparameters to minimize prediction error. The results revealed that the fine-tuning of hyperparameters via the JS algorithm significantly enhanced the precision of the CNN models, resulting in a mean absolute percentage error (MAPE) improvement range of 4.4% to 17.5% compared to the default models. The proposed models can help utilities plan daily power allocation and match supply and demand to maintain a stable power system.

Improving the accuracy of multi-step prediction of building energy consumption based on EEMD-PSO-Informer and long-time series

Accurate multi-step forecasting of building energy consumption is an essential tool for effective planning and plays a vital role in building energy management systems. With the advent of big data, many artificial intelligence techniques require long time series to predict multi-step energy consumption. However, energy consumption data of buildings are often nonlinear and non-stationary in the state as well as considerable period, making model prediction more difficult. In this research, we propose a hybrid algorithm that combines the ensemble empirical modal decomposition (EEMD) and informer, where the parameters of the informer are optimized by the particle swarm optimization algorithm (PSO). At the beginning of our improved method, we use EEMD to break down the raw data into several intrinsic mode functions(IMF) components. Informer is then used to make predictions, and PSO is used to tune hyper-parameters during prediction. In the final step, the final prediction result is obtained by combining the prediction results of all IMF components. The hourly electricity consumption of five buildings in the BDG2 dataset is used to evaluate the effectiveness of the proposed method. Five existing models are compared with this method and evaluated by different performance metrics. From the perspective of five cases, the accuracy rate has increased by up to 78.68% compared with the existing methods. Compared with the original model, the accuracy rate has increased by up to 56.11%.

Nonstationary significant wave height forecasting with a hybrid VMD-CNN model

Significant wave height information is used to measure the intensity of storms and is an important factor in forecasting potential damage in coastal communities, to marine vessels, and to other infrastructure. Under strong nonlinear and nonstationarity conditions, most traditional time series prediction models do not perform well in predicting significant wave heights at a certain location in the sea. This paper proposes a hybrid variational mode decomposition (VMD) and one-dimensional convolutional neural network (1D-CNN) model (VMD-CNN) for nonstationary wave forecasting. The performance of the autoregressive integrated moving average (ARIMA), autoregressive moving average (ARMA), 1D-CNN, and VMD-CNN is studied and verified. In terms of single-step forecasting, the performance of ARMA, ARIMA, 1D-CNN, and VMD-CNN are compared. Moreover, for multistep forecasting, the performance of 1D-CNN and VMD-CNN are compared. Significant wave height data was measured by the WAVERIDER DWR-MKIII wave buoy. The first 90% of the data is used for training, and the rest is used for comparison. The mean square error (MSE), scatter index (SI), coefficient of determination (R2), and time history comparison are used to evaluate the performance of the models. The three metrics indicate that the VMD-CNN model is an effective method for multistep and single-step predictions of nonlinear and nonstationary waves.

An innovative model for electrical load forecasting: A case study in Australia

Short-term load prediction has always played an increasingly important part in power system administration, load dispatch, and energy transfer scheduling. However, how to build a novel model to improve the accuracy of load forecasts is not only an extremely challenging problem but also a concerning problem for the power market. Specifically, the individual model pays no attention to the significance of data selection, data preprocessing, and model optimization. So these models cannot always satisfy the time series forecasting’s requirements. With these above-mentioned ignored factors considered, to enhance prediction accuracy and reduce computation complexity, in this study, a novel and robust method were proposed for multi-step forecasting, which combines the power of data selection, data preprocessing, artificial neural network, rolling mechanism, and artificial intelligence optimization algorithm. Case studies of electricity power data from New South Wales, Australia, are regarded as exemplifications to estimate the performance of the developed novel model. The experimental results demonstrate that the proposed model has significantly increased the accuracy of load prediction in all quarters. As a result, the proposed method not only is simple, but also capable of achieving significant improvement as compared with the other forecasting models, and can be an effective tool for power load forecasting.

Multi-step forecasting of short-term traffic flow based on Intrinsic Pattern Transform

Multi-step forecasting of short-term traffic flow based on Intrinsic Pattern Transform

Multistep Forecasting of Power Flow Based on LSTM Autoencoder: A Study Case in Regional Grid Cluster Proposal

A regional grid cluster proposal is required to tackle power grid complexities and evaluate the impact of decentralized renewable energy generation. However, implementing regional grid clusters poses challenges in power flow forecasting owing to the inherent variability of renewable power generation and diverse power load behavior. Accurate forecasting is vital for monitoring the imported power during peak regional load periods and surplus power generation exported from the studied region. This study addressed the challenge of multistep bidirectional power flow forecasting by proposing an LSTM autoencoder model. During the training stage, the proposed model and baseline models were developed using autotune hyperparameters to fine-tune the models and maximize their performance. The model utilized the last 6 h leading up to the current time (24 steps of 15 min intervals) to predict the power flow 1 h ahead (4 steps of 15 min intervals) from the current time. In the model evaluation stage, the proposed model achieved the lowest RMSE and MAE scores with values of 32.243 MW and 24.154 MW, respectively. In addition, it achieved a good R2 score of 0.93. The evaluation metrics demonstrated that the LSTM autoencoder outperformed the other models for multistep forecasting task in a regional grid cluster proposal.

Artificial Intelligent Power Forecasting for Wind Farm Based on Multi-Source Data Fusion

Wind power forecasting is a typical high-dimensional and multi-step time series prediction problem. Data-driven prediction methods using machine learning show advantages over traditional physical or statistical methods, especially for wind farms with complex meteorological conditions. Thus, effective use of different data sources and data types will help improve power forecasting accuracy. In this paper, a multi-source data fusion method is proposed, which integrates the static information of the wind turbine with observational and forecasting meteorological information together to further improve the power forecasting accuracy. Firstly, the characteristics of each time step are re-characterized by using the self-attention mechanism to integrate the global information of multi-source data, and the Res-CNN network is used to fuse multi-source data to improve the prediction ability of input variables. Secondly, static variable encoding and feature selection are carried out, and the time-varying variables are combined with static variables for collaborative feature selection, so as to effectively eliminate redundant information. A forecasting model based on the Encoder–Decoder framework is constructed with LSTM as the basic unit, and the Add&Norm mechanism is introduced to further enhance the input variable information. In addition, the self-attention mechanism is used to integrate the global time information of the decoded results, and the Time Distributed mechanism is used to carry out multi-step prediction. Our training and testing data are obtained from an operating wind farm in northwestern China. Results show that the proposed method outperforms a classic AI forecasting method such as that using the Seq2Seq+attention model in terms of prediction accuracy, thus providing an effective solution for multi-step forecasting of wind power in wind farms.

A Transformer-based multimodal-learning framework using sky images for ultra-short-term solar irradiance forecasting

The development of solar energy is crucial to combat the global climate change and fossil energy crisis. However, the inherent uncertainty of solar power prevents its large-scale integration into power grids. Although various sky-image-derived modeling methods exist to forecast the variations of solar irradiance, few focus on fully utilizing the coupling correlations between sky images and historical data to improve the forecasting performance. Therefore, a novel multimodal-learning framework is proposed for forecasting global horizontal irradiance (GHI) in the ultra-short-term. First, the historical and empirically estimated clear-sky GHI are encoded by Informer. Then, the ground-based sky images are transformed into optical flow maps, which can be handled by Vision Transformer. Subsequently, a cross-modality attention method is proposed to explore the coupling correlations between the two modalities. Last, a generative decoder is used to implement multi-step forecasting. The experimental results show that the proposed method achieves a normalized root mean square error (NRMSE) of 4.28% in 10-min-ahead forecasting. Several state-of-the-art methods are also used for comparisons. The experimental results show that the proposed method outperforms the benchmark methods and exhibits higher accuracy and robustness in ultra-short-term GHI forecasting.

Predicting the Output of Solar Photovoltaic Panels in the Absence of Weather Data Using Only the Power Output of the Neighbouring Sites

There is an increasing need for capable models in the forecast of the output of solar photovoltaic panels. These models are vital for optimizing the performance and maintenance of PV systems. There is also a shortage of studies on forecasts of the output power of solar photovoltaics sites in the absence of meteorological data. Unlike common methods, this study explores numerous machine learning algorithms for forecasting the output of solar photovoltaic panels in the absence of weather data such as temperature, humidity and wind speed, which are often used when forecasting the output of solar PV panels. The considered models include Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Recurrent Neural Network (RNN) and Transformer. These models were used with the data collected from 50 different solar photo voltaic sites in South Korea, which consist of readings of the output of each of the sites collected at regular intervals. This study focuses on obtaining multistep forecasts for the multi-in multi-out, multi-in uni-out and uni-in uni-out settings. Detailed experimentation was carried out in each of these settings. Finally, for each of these settings and different lookback and forecast lengths, the best models were also identified.

Short-term multi-step wind power forecasting based on spatio-temporal correlations and transformer neural networks

Spatio-temporal wind power forecasting is significant to the stability of electric power systems. However, the accuracy of power forecasting results is easily impaired by the insufficient capacity of sequence modeling and misleading information from distinct wind turbines. In this paper, a novel method is proposed to resolve the mentioned problem. Specifically, the reliability of wind condition knowledge is enhanced by considering the spatial information of surrounding wind turbines. Moreover, two metrics based on the distance and correlation are developed to evaluate the quality of spatial information. To learn sequential dependencies regardless of the distance, wind power modeling is achieved by transformer neural networks based on the multi-head attention mechanism. Furthermore, experiments are conducted to assess the performance of the proposed method with real-world measurements. Results show that the proposed method outperforms several baseline and state-of-the-art approaches, and the superiority is particularly prominent with large steps. In two experiments, the average values of mean absolute error of forecasting results generated by the proposed method are only 0.0914 and 0.0911, respectively, which is significantly better than other approaches. With accurate results of short-term multi-step forecasting, this work makes contributions to the effective utilization of wind energy resources.

Cutting state estimation and time series prediction using deep learning for Cutter Suction Dredger

With the increasing environmental protection standard, the Cutter Suction Dredger (CSD) construction is required to guarantee precise dredging in some special areas. Specifically, dredging the side slope of channeling, which requires the operator to accurately control the CSD. However, the underwater cutting environment of CSD construction is complicated and the soil kinds are sometimes spatially different, which makes it extremely difficult for the operator to accurately control the CSD. Therefore, we propose a state-of-the-art time series prediction model based on CNN-LSTM Encode-Decode, which can predict the cutting status of the cutter in advance and also ahead estimate the soil kinds to be cut. First, we establish the relationship between the cutting torque and the cutter motor power (CMP) and rotation speed (CMRS). By ahead forecasting the CMP and CMRS, the cutting torque at the next moment can be obtained. Then, the experienced operator can estimate the kind of underwater cutting soil by torque. In addition, we also compare the ahead one-step forecasting and ahead multi-step forecasting of this method with other models and conduct an experiment to verify the performance. The comparison results demonstrate that our approach significantly outperforms, which can help the operator's ahead perception of the underwater cutter status and ensure precise dredging.

Effective RNN-Based Forecasting Methodology Design for Improving Short-Term Power Load Forecasts: Application to Large-Scale Power-Grid Time Series

This article introduces a carefully-engineered forecasting methodology for day-ahead electric power load forecasts evaluated using the European Network of Transmission System Operators for Electricity (ENTSO-E). Two steps were employed to configure the desired forecasting methodology: First, a straightforward processing pipeline is proposed to enable systematic preprocessing of raw multivariate time-discrete power data extracted from the ENTSO-E repository, including a stride-based sliding window approach to generate time series-based batches ready for the supervised learning procedure. Second, the lightweight type of recurrent neural network method, namely gated recurrent units (GRU), is selected and carefully calibrated to yield accurate multi-step forecasts, which was trained using the preprocessed multivariate time series data to render day-ahead power load forecasts. The forecasting estimates generated by the proposed GRU model are evaluated using a set of regression-based metrics to assess the models’ precisions. The empirical results show that the proposed forecasting methodology yields outstanding day-ahead power load forecasting performance regarding the enterprise-class measured data compared to a statistical model, namely autoregressive integrated moving average with exogenous variables (ARIMAX), as well as the actual day-ahead forecasts generated by the ENTSO-E platform.