Hourly Forecast Research Articles

A model coupling current non-adjustable, coming adjustable and remaining stages for daily generation scheduling of a wind-solar-hydro complementary system

To formulate the daily generation scheduling of a wind-solar-hydro complementary system (WSHCS), the hourly forecasts of the reservoir inflow, wind speed, and sunlight intensity within a day and their longer prediction can be used to assess the benefit of the day-ahead and remaining stages, respectively. Thus, the current two-stage model coupling these forecasts can efficiently maximize the total benefits of day-ahead and remaining stages. However, the two-stage model does not consider the significant difference between non-adjustable and adjustable periods for hydropower. Namely, the hydro unit commitment should be decided in the current stage (non-adjustable period), and it can be adjusted in the coming stage (adjustable period). To address this issue, a three-stage model, including the day-ahead non-adjustable stage, the day-ahead adjustable stage, and the remaining period stage, is proposed to formulate a daily generation scheduling of the WSHCS. The Guandi wind-solar-hydro hybrid power plant on China's Yalong River is selected as a case study. Results show that compared with the two-stage model, the proposed three-stage operation model can increase the average energy production of the WSHCS by 2.23 GWh for typical days, and the average power curtailment rate reduces by 0.26%. A three-layer nested approach can simplify from 576 to 26 decision variables for the high-dimensional three-stage model, and improve the computational efficiency. Thus, the proposed method could guide the daily generation scheduling of a WSHCS.

AI-based air quality PM2.5 forecasting models for developing countries: A case study of Ho Chi Minh City, Vietnam

Outdoor air pollution damages the climate and causes many diseases, including cardiovascular diseases, respiratory infections, and lung damage. In particular, Particulate Matter (PM2.5) is considered a hazardous air pollutant to human health. Accurate hourly forecasting of PM2.5 concentrations is thus of significant importance for public health, helping the citizens to plan the measures to alleviate the harmful effects of air pollution on health.This study analyses and discusses the temporal characteristics of PM2.5 at different locations in Ho Chi Minh City (HCMC), Vietnam - an economic center and a megacity in a developing country with a population of 8.99 million people. We developed several AI-based one-shot multi-step PM2.5 forecasting models, with both an hourly forecast granularity (1 h to 24 h) and a 24-h rolling mean. These Machine Learning algorithms include Stochastic Gradient Descent Regressor, hybrid 1D CNN-LSTM, eXtreme Gradient Boosting Regressor, and Prophet. We collected the data from six monitoring stations installed by the HealthyAir project partners at different locations in HCMC, including traffic, residential and industrial areas in the city. In addition, we developed a suitable model training protocol using data from a short period to address the non-stationarity of PM2.5 time series. Our proposed PM2.5 forecasting models achieve state-of-the-art accuracy and will be deployed in our HealthyAir mobile app to warn HCMC citizens of air pollution issues in the city.

A significant wave height prediction method based on deep learning combining the correlation between wind and wind waves

Accurate wave height prediction is significant in ports, energy, fisheries, and other offshore operations. In this study, a regional significant wave height prediction model with a high spatial and temporal resolution is proposed based on the ConvLSTM algorithm. The model learns the intrinsic correlations of the data generated by the numerical model, making it possible to combine the correlations between wind and wind waves to improve the predictions. In addition, this study also optimizes the long-term prediction ability of the model through the proposed Mask method and Replace mechanism. The experimental results show that the introduction of the wind field can significantly improve the significant wave height prediction results. The research on the prediction effect of the entire study area and two separate stations shows that the prediction performance of the proposed model is better than the existing methods. The model makes full use of the physical correlation between wind and wind waves, and the validity is up to 24 hours. The 24-hour forecast R² reached 0.69.

A generic methodology to efficiently integrate weather information in short-term Photovoltaic generation forecasting models

The power output of Photovoltaic (PV) plants is weather-dependent, which results in inherent uncertainties about future production. This raises technical challenges for grid operators, especially in power systems with high PV penetration, and also financial losses when PV generation is traded on electricity markets. Accurate forecasts for the next hours or days contribute to alleviating these impacts. The literature features a plethora of forecasting models, among which outstanding approaches combine heterogeneous sources of inputs like measurements, weather forecasts and satellite images. The integration of such inputs into forecast models can take two forms: either as explanatory features, or as state features that condition the model training through a local regression approach. With the latter, physics-based information can be included within statistical regression tools to derive optimised models w.r.t. weather input. These models are then extended to integrate spatio-temporal information from satellite observations. We investigate these approaches with the objective of deriving the mathematical foundations of a generic methodology to integrate weather information into PV forecasting models. The paper assesses the influence of weather information integration strategies on forecasting performances for two state-of-the-art short-term forecasting models, belonging respectively to linear and non-linear families. Lastly, general guidelines for forecasters are derived regarding the procedure to follow when dealing with several sources of information. Evaluations are performed on real-world datasets composed of nine PV plants.

On Hourly Forecasting Heating Energy Consumption of HVAC with Recurrent Neural Networks

Optimizing the heating, ventilation, and air conditioning (HVAC) system to minimize district heating usage in large groups of managed buildings is of the utmost important, and it requires a machine learning (ML) model to predict the energy consumption. An industrial use case to reach large building groups is restricted to using normal operational data in the modeling, and this is one reason for the low utilization of ML in HVAC optimization. We present a methodology to select the best-fitting ML model on the basis of both Bayesian optimization of black-box models for defining hyperparameters and a fivefold cross-validation for the assessment of each model’s predictive performance. The methodology was tested in one case study using normal operational data, and the model was applied to analyze the energy savings in two different practical scenarios. The software for the modeling is published on GitHub. The results were promising in terms of predicting the energy consumption, and one of the scenarios also showed energy saving potential. According to our research, the GitHub software for the modeling is a good candidate for predicting the energy consumption in large building groups, but further research is needed to explore its scalability for several buildings.

Calibrating the CAMS European multi-model air quality forecasts for regional air pollution monitoring

The CAMS air quality multi-model forecasts have been assessed and calibrated for PM10, PM2.5, O3, NO2, and CO against observations collected by the Regional Monitoring Network of the Liguria region (northwestern Italy) in the years 2019 and 2020. The calibration strategy used in the present work has its roots in the well-established Ensemble Model Output Statistics (EMOS) through which a raw ensemble forecast can be accurately transformed into a predictive probability density function, with a simultaneous correction of biases and dispersion errors. The strategy also provides a calibrated forecast of model uncertainties. As a result of our analysis, the key role of pollutant real-time observations to be ingested in the calibration strategy clearly emerge especially in the shorter look-ahead forecast hours. Our dynamic calibration strategy turns out to be superior with respect to its analogous where real-time data are not taken into account. The best calibration strategy we have identified makes the CAMS multi-model forecast system more reliable than other raw air quality models running at higher spatial resolution which exploit more detailed information from inventory emission. We expect positive impacts of our research for identifying and set up reliable and economic air pollution early warning systems.

STG-OceanWaveNet: Spatio-temporal geographic information guided ocean wave prediction network

This study proposes a spatio-temporal geographical information-guided neural network to predict multi-step ahead space-time series of ocean waves. The network is designed to learn the ocean wave dynamics from external atmospheric forcing and internal wave processes. It also captures the nonlinear relationships in multiple input and at the spatial and temporal levels and shares their dependencies. The model learns these dependencies through a convoluted encoded latent feature, while a decoder predicts the space-time series of ocean waves from the latent representations. The model is trained on 35 years of a state-of-the-art global reanalysis dataset produced at 1-hour temporal and 25 km spatial resolutions around the Korean Peninsula. It is evaluated by predicting the same resolution’s multi-step ahead space-time series of ocean waves for a 48-hour forecast lead time for the 5 years not used for training. We conducted an ablation experiment to determine the optimal model architecture, input variable, and training period. The prediction results are compared and analyzed with the in-situ ocean wave measurements at the 18 observation stations. We consider the prediction results according to the start time of prediction with the in-situ measurements and hindcast results to examine the performance on the high waves that caused wave-induced disaster.

Day-ahead Electricity Price Forecasting using Long-short Term Memory Recurrent Neural Network

The availability of accurate day-ahead electricity price forecasts is very important for electricity market participants and it is an essential challenge to accurately forecast the electricity price. Therefore, this study proposes an efficient method suitable for electricity price forecasting (EPF) and processing time-series data from the Bulgarian day-ahead market based on a long-short term memory (LSTM) recurrent neural network model. The LSTM model is used to forecast the day-ahead electricity price for the Bulgarian day-ahead market. As inputs to the model are used historical hourly prices for the period between 20.01.2016 and 05.03.2022. The output is the electricity price forecasts for hours and days ahead. The future values of prices are forecasted recursively. LSTM can model temporal dependencies in larger Time Series set horizons without forgetting the short-term patterns. LSTM networks are composed of units that are called LSTM memory cells and these cells contain some gates that process the inputs. Since electricity price is affected by various seasonal effects, the model is trained for several years. The effectiveness of the proposed method is verified using real market data.

Analysis of methods of electricity consumption forecasting for a coal industry enterprise

The paper considers a forecast model of electricity consumption of a coal industry enterprise based on three forecast methods, namely the wavelet transform, the vector method, and the recurrent neutral network. For preprocessing the data for forecasting by vector and recurrent methods, the Singular Spectrum Analysis method was chosen. The structure of the model allows taking into account individual features of the operating cycle of the production process and smoothing the noise components and outliers caused by these features. The results of a short-term hourly forecast for one day ahead are presented with the comparison of the obtained values. The results of short-term electricity consumption forecast were verified based on the actual data of the coal industry enterprise in order to assess the adequacy of the model to the actual values.

Methodology Based on Artificial Neural Networks for Hourly Forecasting of PV Plants Generation

We propose a methodology for the hourly forecast of the photovoltaic (PV) power of two plants (6.03 kWp and 7.37 kWp) installed in the metropolitan region of Fortaleza - CE. The methodology uses two Artificial Neural Networks (ANN) to predict time series: Perceptron type with Multiple Layers (MLP) and radial base functions (RBF), trained with historical data of hourly PV power collected during the year 2020 in the locations under study. System performance meters are applied (correlation coefficient - R, Nash-Sutcliffe efficiency - NSE and relative trend - VR). The data evaluated in each plant are treated using MLP and RBF networks, as well as the Persistence method, seeking to increase the study reliability. ANN results indicate potential to learn the behavior of the plants, with R above 80%, VR close to zero and NSE above 0.50 in two of the applications. In this specific case, despite being similar networks, MLP shows a higher accuracy than RBF.

Comparison of the Performance of the Observation-Based Hybrid EDMF and EDMF-TKE PBL Schemes in 2020 Tropical Cyclone Forecasts from the Global-Nested Hurricane Analysis and Forecast System

Abstract Better representation of the planetary boundary layer (PBL) in numerical models is one of the keys to improving forecasts of TC structure and intensity, including rapid intensification. To meet this goal, our recent work has used observations to improve the eddy-diffusivity mass flux with prognostic turbulent kinetic energy (EDMF-TKE) PBL scheme in the Hurricane Analysis and Forecast System (HAFS). This study builds on that work by comparing a modified version of EDMF-TKE (MEDMF-TKE) with the hybrid EDMF scheme based on a K-profile method (HEDMF-KP) in the 2020 HAFS-globalnest model. Verification statistics based on 101 cases in the 2020 season demonstrate that MEDMF-TKE improves track forecasts, with a reduction in a large right bias seen in HEDMF-KP forecasts. The comparison of intensity performance is mixed, but the magnitude of low bias at early forecast hours is reduced with the use of the MEDMF-TKE scheme, which produces a wider range of TC intensities. Wind radii forecasts, particularly the radius of maximum wind speed (RMW), are also improved with the MEDMF-TKE scheme. Composites of TC inner-core structure in and above the PBL highlight and explain differences between the two sets of forecasts, with MEDMF-TKE having a stronger and shallower inflow layer, stronger eyewall vertical velocity, and more moisture in the eyewall region. A case study of Hurricane Laura shows that MEDMF-TKE better represented the subtropical ridge and thus the motion of the TC. Finally, analysis of Hurricane Delta through a tangential wind budget highlights how and why MEDMF-TKE leads to faster spinup of the vortex and a better prediction of rapid intensification.

Application of Acoustic tomographic data to the short-term streamflow forecasting using data-driven methods and discrete wavelet transform

High-accurate streamflow data obtained from the Fluvial Acoustic Tomography (FAT) increase forecasting accuracy, especially in the case of short-term forecasting. The reason is due to the better temporal resolution and higher measurement frequency than the traditional Rating-Curve (RC) method. In this study, Multilayer GMDH Algorithm (MGA) and Combinatorial GMDH Algorithm (CGA) were used to forecast the short-term streamflow. These models, as well as their combinations with the discrete wavelet transform (DWT), were applied to the 6-month streamflow datasets obtained from the FAT and RC method, and the hourly forecasts of streamflow were done with 1 to 48 h ahead lead times. The results showed that the MGA model has a better performance than the CGA model. It was also observed that combining of the MGA model with DWT increased the forecasting accuracy. However, based on the concept of Occam’s razor and due to the insignificant difference between MGA and DWT-MGA models, the simpler one, i.e., MGA, was chosen as the optimal model. The comparison of forecasting results using the FAT and RC revealed that in the case of applying FAT dataset, the Nash-Sutcliffe (NSE) rate ranged from 0.99 to 0.73 for 1-hour ahead to 48-hour ahead forecasts, whereas, it adversely changed from 0.98 to 0.06 for the RC dataset. This study demonstrates the capability of FAT data in providing reliable and accurate hourly streamflow forecasting.

An FSO-based Optimization Framework for Improved Observation Performance: Theoretical Formulation and Experiments with NAVDAS-AR/NAVGEM

Abstract The forecast sensitivity to observations (FSO) is embedded into a new optimization framework for improving the observation performance in atmospheric data assimilation. Key ingredients are introduced as follows: the innovation-weight parametrization of the analysis equation, the FSO-based evaluation of the forecast error gradient to parameters, a line search approach to optimization, and an efficient mechanism for step length specification. This methodology is tested in preliminary numerical experiments with the Naval Research Laboratory Atmospheric Variational Data Assimilation System-Accelerated Representer (NAVDAS-AR) and the U.S. Navy’s Global Environmental Model (NAVGEM) at a T425L60 resolution. The experimental setup relies on a verification state produced by the European Centre for Medium-Range Weather Forecasts (ECMWF) to estimate the analysis and short-range forecast errors. Parameter tuning is implemented in a training stage valid for April 1-14 of 2018 and aimed at improving the use of assimilated observations in reducing the initial-condition errors. Assessment is carried out for April 15 – May 31 of 2018 to investigate the performance of the weighted assimilation system in reducing the errors in analyses and 24-hour model forecasts. In average, as compared with the control run and verified against the ECMWF analyses, the weighted approach provided 17.4% reduction in analysis errors and 3.1% reduction in 24-hour forecast errors, measured in a dry total energy norm. Observation impacts are calculated to assess the use of various observation types in reducing the analysis and forecast errors. In particular, assimilation of satellite wind data is significantly improved through the innovation-weighting procedure.

Effects of Learning Rates and Optimization Algorithms on Forecasting Accuracy of Hourly Typhoon Rainfall: Experiments With Convolutional Neural Network

AbstractThe current study used seven optimization algorithms such as stochastic gradient descent (SGD), root mean square propagation (RMSprop), adaptive grad (AdaGrad), adaptive delta (AdaDelta), adaptive moment estimation (Adam), adaptive maximum (Adamax), and Nesterov‐accelerated adaptive moment estimation (Nadam) and eight learning rates (1, 0.1, 0.01, 0.001, 0.0001, le‐05, le‐06, and le‐07) to investigate the effects of these learning rates and optimizers on the forecasting performance of the convolutional neural network (CNN) model to forecast hourly typhoon rainfall. The model was developed using antecedent hourly typhoon rainfall within a 500 km radius from each typhoon center. Results showed that too‐large and too‐small learning rates would result in the inability of the model to learn anything to forecast hourly typhoon rainfall. The CNN model showed the best performance for learning rates of 0.1, 0.01, and 0.001 to forecast hourly typhoon rainfall. For long‐lead‐time forecasting (1–6 hr), the CNN model with SGD, RMSprop, AdaGrad, AdaDelta, Adam, Adamax, Nadam optimizers and learning rates of 0.1, 0.01, and 0.001 showed more accurate forecasts than the existing models. Therefore, this study recommends that future work may consider the CNN model as an alternative to the existing model for disaster warning systems.

Using a Cost-Effective Approach to Increase Background Ensemble Member Size within the GSI-Based EnVar System for Improved Radar Analyses and Forecasts of Convective Systems

Abstract A valid time shifting (VTS) method is explored for the GSI-based ensemble variational (EnVar) system modified to directly assimilate radar reflectivity at convective scales. VTS is a cost-efficient method to increase ensemble size by including subensembles before and after the central analysis time. Additionally, VTS addresses common time and phase model error uncertainties within the ensemble. VTS is examined here for assimilating radar reflectivity in a continuous hourly analysis system for a case study of 1–2 May 2019. The VTS implementation is compared against a 36-member control experiment (ENS-36), to increase ensemble size (3 × 36 VTS), and as a cost-savings method (3 × 12 VTS), with time-shifting intervals τ between 15 and 120 min. The 3 × 36 VTS experiments increased the ensemble spread, with largest subjective benefits in early cycle analyses during convective development. The 3 × 12 VTS experiments captured analysis with similar accuracy as ENS-36 by the third hourly analysis. Control forecasts launched from hourly EnVar analyses show significant skill increases in 1-h precipitation over ENS-36 out to hour 12 for 3 × 36 VTS experiments, subjectively attributable to more accurate placement of the convective line. For 3 × 12 VTS, experiments with τ ≥ 60 min met and exceeded the skill of ENS-36 out to forecast hour 15, with VTS-3 × 12τ90 maximizing skill. Sensitivity results demonstrate preference to τ = 30–60 min for 3 × 36 VTS and 60–120 min for 3 × 12 VTS. The best 3 × 36 VTS experiments add a computational cost of 45%–67%, compared to the near tripling of costs when directly increasing ensemble size, while best 3 × 12 VTS experiments save about 24%–41% costs over ENS-36. Significance Statement The purpose of this work is to study a valid time shifting method to improve the prediction of severe convective storm systems over the continental United States. This method improves ensemble-based radar reflectivity analyses by including ensemble member information at times before and after the analysis time, thereby increasing the ensemble size at just a fractional added computational cost. The results show the method can boost the accuracy of high-resolution convection prediction out to at least 12 h. This case study motivates future systematic testing in a real-time setting and potential implementation to enhance the U.S. operational ensemble-based convection-allowing forecast model and data assimilation system.

Vertical power flow forecast with LSTMs using regular training update strategies

The strong growth of renewable energy sources as well as the increasing amount of volatile energy consumption is leading to major challenges in the electrical grid. In order to ensure safety and reliability in the electricity grid, a high quality of power flow forecasts for the next few hours are needed. In this paper we investigate forecasts of the vertical power flow at transformer between the medium and high voltage grid. Forecasting the vertical power flow is challenging due to constantly changing characteristics of the power flow at the transformer. We present a novel approach to deal with these challenges. For the multi step time series forecasts a Long-Short Term Memory (LSTM) is used. In our presented approach an update process where the model is retrained regularly is investigated and compared to baseline models. The model is retrained as soon as a sufficient amount of new measurements are available. We show that this retraining mostly captures changes in the characteristic of the transformer that the model has not yet seen in the past and therefore cannot be predicted by the model without an update process. To give more weight to recent data, we examined different strategies in terms of the number of epochs and the learning rate. We show that our new approach significantly outperforms the investigated baseline models. On average, we achieved an improvement of about 8% with the regular update process compared to the approach without update process.

Air Quality Forecasts Improved by Combining Data Assimilation and Machine Learning With Satellite AOD

AbstractSatellite aerosol optical depth (AOD) data assimilation (DA) using numerical air quality forecast models has shown a limited improvement due to large uncertainties in the AOD observation operator. This study employed a machine learning (ML) algorithm to estimate the ground‐level particulate matter (PM) from the Geostationary Ocean Color Imager (GOCI) AOD through the random forest with high accuracy. Analysis fields were subsequently produced by applying PM estimations to the Weather Research and Forecasting‐Chemistry/three‐dimensional variational DA system. Initialization of the model with the new analysis remarkably reduced the analysis error and increased the forecast skill. The PM10 prediction showed significant benefits for up to 24 forecast hours, whereas PM2.5 prediction was improved for up to six forecast hours. Considering a broad spatial coverage by satellites, the synergistic use of DA and ML can maximize the effectiveness of satellite DA for air quality forecasts at the ground.

Beach-level 24-hour forecasts of Florida red tide-induced respiratory irritation

An accurate forecast of the red tide respiratory irritation level would improve the lives of many people living in areas affected by algal blooms. Using a decades-long database of daily beach conditions, two conceptually different models to forecast the respiratory irritation risk level one day ahead of time are trained. One model is wind-based, using the current days’ respiratory level and the predicted wind direction of the following day. The other model is a probabilistic self-exciting Hawkes process model. Both models are trained on beaches in Florida during 2011--2017 and applied to the red tide bloom during 2018-2019. For beaches where there is enough historical data to develop a model, the model which performs best depends on the beach. The wind-based model is the most accurate at half the beaches, correctly predicting the respiratory risk level on average about 84% of the time. The Hawkes model is the most accurate (81% accuracy) at nearly all of the remaining beaches.

Air-pollution prediction in smart city, deep learning approach

Over the past few decades, due to human activities, industrialization, and urbanization, air pollution has become a life-threatening factor in many countries around the world. Among air pollutants, Particulate Matter with a diameter of less than 2.5 mu m (PM_{2.5}) is a serious health problem. It causes various illnesses such as respiratory tract and cardiovascular diseases. Hence, it is necessary to accurately predict the PM_{2.5} concentrations in order to prevent the citizens from the dangerous impact of air pollution beforehand. The variation of PM_{2.5} depends on a variety of factors, such as meteorology and the concentration of other pollutants in urban areas. In this paper, we implemented a deep learning solution to predict the hourly forecast of PM_{2.5} concentration in Beijing, China, based on CNN-LSTM, with a spatial-temporal feature by combining historical data of pollutants, meteorological data, and PM_{2.5} concentration in the adjacent stations. We examined the difference in performances among Deep learning algorithms such as LSTM, Bi-LSTM, GRU, Bi-GRU, CNN, and a hybrid CNN-LSTM model. Experimental results indicate that our method “hybrid CNN-LSTM multivariate” enables more accurate predictions than all the listed traditional models and performs better in predictive performance.

Improving the Computational Efficiency of the Global SL-AV Numerical Weather Prediction Model

The recent works on improving the efficiency of the Russian SL-AV global numerical weather prediction model both for medium- and long-range forecasts are described. The algorithmic improvements of SL-AV dynamical core, implementation of parallel I/O and several code optimizations are presented. We investigate the impact of single precision computations in some parts of the code on present climate simulations. As a result of efforts described in this article, we are now able to compute a 24-hour forecast for the model version having about 10 km horizontal resolution and 104 vertical levels in 13 min using 2916 processor cores of Cray XC40 system. This timing allows multiple experiments for tuning this new model and fits the requirements for operational weather forecast. The single long-range forecast with low-resolution SL-AV version now takes just 89 minutes instead of 111. We have also verified that the partial utilization of single precision computations produces approximately the same model climate as the previous version with fully double precision computations.