Online short-term solar power forecasting

This paper describes a new approach to online forecasting of power production from PV systems. The method is suited to online forecasting in many applications and in this paper it is used to predict hourly values of solar power for horizons of up to 36 h. The data used is 15-min observations of solar power from 21 PV systems located on rooftops in a small village in Denmark. The suggested method is a two-stage method where first a statistical normalization of the solar power is obtained using a clear sky model. The clear sky model is found using statistical smoothing techniques. Then forecasts of the normalized solar power are calculated using adaptive linear time series models. Both autoregressive (AR) and AR with exogenous input (ARX) models are evaluated, where the latter takes numerical weather predictions (NWPs) as input. The results indicate that for forecasts up to 2 h ahead the most important input is the available observations of solar power, while for longer horizons NWPs are the most important input. A root mean square error improvement of around 35% is achieved by the ARX model compared to a proposed reference model.     

Modeling and forecasting the mean hourly wind speed time series using GMDH-based abductive networks

Wind speed forecasts are important for the operation and maintenance of wind farms and their profitable integration into power grids, as well as many important applications in shipping, aviation, and the environment. Modern machine learning techniques including neural networks have been used for this purpose, but it has proved hard to make significant improvements on the performance of the simple persistence model. As an alternative approach, we propose here the use of abductive networks, which offer the advantages of simplified and more automated model synthesis and transparent analytical input–output models. Various abductive models for predicting the mean hourly wind speed 1 h ahead have been developed using wind speed data at Dhahran, Saudi Arabia during the month of May over the years 1994–2005. The models were evaluated on the data for May 2006. Models described include a single generic model to forecast next-hour speed from the previous 24 hourly measurements and an hour index, which give an overall mean absolute error (MAE) of 0.85 m/s and a correlation coefficient of 0.83 between actual and predicted values. The model achieves an improvement of 8.2% reduction in MAE compared to hourly persistence. The above model was used iteratively to forecast the hourly wind speed 6 h and 24 h ahead at the end of a given day, with MAEs of 1.20 m/s and 1.42 m/s which are lower than forecasting errors based on day-to-day persistence by 14.6% and 13.7%. Relative improvements on persistence exceed those reported for several machine learning approaches reported in the literature.     

Predicting and presenting plausible future scenarios of wind power production from numerical weather prediction systems: a qualitative ex ante evaluation for decision making

This paper is concerned with evaluating techniques to forecast plausible future scenarios in wind power production for up to 48 h ahead, where the term scenario refers to a coherent chronological prediction including the timing, rapidity and size of large changes. Such predictions are of great interest in power systems with high regional wind penetration where a large rapid change in wind power may pose a threat to power system security. Numerous studies have evaluated wind power forecasting methods on ex post statistical measures of forecast accuracy such as root mean square error. Other work has assessed the forecast value by simulating automated decision making for bidding wind generation into particular electricity markets, and in some cases, the ex ante value of a perfect forecast has been assessed. The future, however, will always be uncertain, and decision making always takes place in an ex ante context. This paper discusses how numerical weather prediction (NWP) systems forecasts are produced, with a particular focus on uncertainty and how forecasters might visually present plausible future scenarios for wind power to electricity industry decision makers. It is difficult to quantify the ex ante value of visual wind power forecast information to the complex decision‐making process involved. Consequently, this paper explores qualitative assessments of ex ante value by proposing six desirable attributes for the techniques and the presentation of NWP forecasts to decision makers. It uses these attributes to assess four such methodologies, which include NWP ensemble methods and the recently introduced NWP spatial field approach. Copyright © 2011 John Wiley & Sons, Ltd.     

A new recursive neural network algorithm to forecast electricity price for PJM day‐ahead market

This paper evaluates the usefulness of publicly available electricity market information in predicting the hourly prices in the PJM day‐ahead electricity market using recursive neural network (RNN) technique, which is based on similar days (SD) approach. RNN is a multi‐step approach based on one output node, which uses the previous prediction as input for the subsequent forecasts. Comparison of forecasting performance of the proposed RNN model is done with respect to SD method and other literatures. To evaluate the accuracy of the proposed RNN approach in forecasting short‐term electricity prices, different criteria are used. Mean absolute percentage error, mean absolute error and forecast mean square error (FMSE) of reasonably small values were obtained for the PJM data, which has correlation coefficient of determination (R²) of 0.7758 between load and electricity price. Error variance, one of the important performance criteria, is also calculated in order to measure robustness of the proposed RNN model. The numerical results obtained through the simulation to forecast next 24 and 72 h electricity prices show that the forecasts generated by the proposed RNN model are significantly accurate and efficient, which confirm that the proposed algorithm performs well for short‐term price forecasting. Copyright © 2009 John Wiley & Sons, Ltd.     

Forecast of hourly average wind speed with ARMA models in Navarre (Spain)

In this article we have used the ARMA (autoregressive moving average process) and persistence models to predict the hourly average wind speed up to 10 h in advance. In order to adjust the time series to the ARMA models, it has been necessary to carry out their transformation and standardization, given the non-Gaussian nature of the hourly wind speed distribution and the non-stationary nature of its daily evolution. In order to avoid seasonality problems we have adjusted a different model to each calendar month. The study expands to five locations with different topographic characteristics and to nine years. It has been proven that the transformation and standardization of the original series allow the use of ARMA models and these behave significantly better in the forecast than the persistence model, especially in the longer-term forecasts. When the acceptable RMSE (root mean square error) in the forecast is limited to 1.5 m/s, the models are only valid in the short term.     

Detecting,categorizing and forecasting large ramps in wind farm power output using meteorological observations and WPPT

The Wind Power Prediction Tool (WPPT) has been installed in Australia for the first time, to forecast the power output from the 65MW Roaring 40s Renewable Energy P/L Woolnorth Bluff Point wind farm. This article analyses the general performance of WPPT as well as its performance during large ramps (swings) in power output. In addition to this, detected large ramps are studied in detail and categorized. WPPT combines wind speed and direction forecasts from the Australian Bureau of Meteorology regional numerical weather prediction model, MesoLAPS, with real‐time wind power observations to make hourly forecasts of the wind farm power output. The general performances of MesoLAPS and WPPT are evaluated over 1 year using the root mean square error (RMSE). The errors are significantly lower than for basic benchmark forecasts but higher than for many other WPPT installations, where the site conditions are not as complicated as Woolnorth Bluff Point. Large ramps are considered critical events for a wind power forecast for energy trading as well as managing power system security. A methodology is developed to detect large ramp events in the wind farm power data. Forty‐one large ramp events are detected over 1 year and these are categorized according to their predictability by MesoLAPS, the mechanical behaviour of the wind turbine, the power change observed on the grid and the source weather event. During these events, MesoLAPS and WPPT are found to give an RMSE only roughly equivalent to just predicting the mean (climatology forecast). Copyright © 2007 John Wiley & Sons, Ltd.     

Monthly mean hourly global solar radiation estimation

In this paper, selected empirical models were used to estimate the monthly mean hourly global solar radiation from the daily global radiation at three sites in the east coast of Malaysia. The purpose is to determine the most accurate model to be used for estimating the monthly mean hourly global solar radiation in these sites. The hourly global solar radiation data used for the validation of selected models were obtained from the Malaysian Meteorology Department and University Malaysia Terengganu Renewable Energy Station. In order to indicate the performance of the models, the statistical test methods of the normalized mean bias error, normalized root mean square error, correlation coefficient and t-statistical test were used. The monthly mean hourly global solar radiation values were calculated by using six models and the results were compared with corresponding measured data. All the models fit the data adequately and can be used to estimate the monthly mean hourly global solar radiation. This study finds that the Collares-Pereira and Rabl model performed better than the other models. Therefore the Collares-Pereira and Rabl model is recommended to estimate the monthly mean hourly global radiations for the east coast of Malaysia with humid tropical climate and in elsewhere with similar climatic conditions.     

Developing an improved global solar radiation map for Zimbabwe through correlating long-term ground- and satellite-based monthly clearness index values

Reliable knowledge of the spatio-temporal distribution of solar radiation is required for the informed design and deployment planning of solar energy delivery systems. In this paper an improved global solar radiation map for Zimbabwe is developed by merging ground-measured radiation data from a sparsely distributed station network, with less accurate satellite-measured data which have an almost continuous spatial coverage. Monthly clearness index values derived from ground-measured global radiation are correlated with those derived from satellite data to obtain a model for calibrating satellite-measured data at a specified grid interval. Two multiplicative factors are to then used to further correct the generated data; CF_m to cater for the in-exactness of the regression fit and the other, IBCF to cater for the interpolation error. Contour maps of global solar radiation are then constructed using interpolation by the geo-statistical method of ordinary kriging. The accuracy of the maps in predicting observed (ground-measured) values was tested by evaluating error statistics; relative bias error (rBE), relative mean bias error (rMBE) and normalized root mean square error (NRMSE) in a “leave-one-out” cross-validation analysis. Results indicate that the maximum normalized root mean square error was 0.028 (about 3%), a significant improvement when compared to an earlier map, the H–G map with a normalized root mean square error (NRMSE) of 0.097.     

基于CA-SE-GA-BP的光伏发电功率预测

针对光伏发电功率预测精度低的问题,以澳大利亚爱丽丝泉地区某200kW的光伏电站为例,选用遗传算法(GA)优化BP神经网络,采用相关性分析法(CA)确定太阳辐照度、温度、湿度为影响光伏发电功率的主要因子,结合经样本熵(SE)量化的天气类型作为模型输入量,提出CA-SE-GA-BP神经网络的光伏发电功率预测模型。结果表明,多云天气下CA-SE-GA-BP神经网络均方根误差、平均绝对百分比误差分别为4.48%、2.27%,晴天、雾霾、雨天三种天气类型下的预测误差也基本上不超过10%,相较于SE-GA-BP、CA-GA-BP、GA-BP神经网络,CA-SE-GA-BP神经网络预测误差降低,为解决光伏系统发电功率预测提供了一种高效准确可行的方法。     

Hybrid intra-hour DNI forecasts with sky image processing enhanced by stochastic learning

We propose novel smart forecasting models for Direct Normal Irradiance (DNI) that combine sky image processing with Artificial Neural Network (ANN) optimization schemes. The forecasting models, which were developed for over 6 months of intra-minute imaging and irradiance measurements, are used to predict 1 min average DNI for specific time horizons of 5 and 10 min. We discuss optimal models for low and high DNI variability seasons. The different methods used to develop these season-specific models consist of sky image processing, deterministic and ANN forecasting models, a genetic algorithm (GA) overseeing model optimization and two alternative methods for training and validation. The validation process is carried over by the Cross Validation Method (CVM) and by a randomized training and validation set method (RTM). The forecast performance for each solar variability season is evaluated, and the models with the best forecasting skill for each season are selected to build a hybrid model that exhibits optimal performance for all seasons. An independent testing set is used to assess the performance of all forecasting models. Performance is assessed in terms of common error statistics (mean bias and root mean square error), but also in terms of forecasting skill over persistence. The hybrid forecast models proposed in this work achieve statistically robust forecasting skills in excess of 20% over persistence for both 5 and 10 min ahead forecasts, respectively.     

Spatial insolation models for photovoltaic energy in Canada

Spatial models of global insolation and photovoltaic electricity generation potential for Canada were developed. The main objective was to provide Canadians with an easily accessible, reliable tool for rapidly estimating the monthly and yearly electricity production potential of grid-connected photovoltaic systems anywhere in the country, and for assessing the dependence of production on location, time of year and array orientation. Monthly mean daily insolation data from 144 meteorological stations across Canada were used, along with data from an additional eight stations in Alaska to improve the models in that region. Several photovoltaic array orientations were considered, including South-facing arrays with latitude and vertical tilts and a sun-tracking orientation. Partial thin plate smoothing splines as implemented in ANUSPLIN were used to generate the spatial insolation models. The models were based on geographic position and a transform of monthly mean precipitation, the latter variable being a surrogate for cloudiness which affects surface insolation. Photovoltaic electricity generation (in kW h per kilowatt of photovoltaic installed power capacity) was estimated for each month and for the entire year from the insolation models by assuming international standard values for the performance ratio of photovoltaic systems. The yearly average root mean square predictive error (RTGCV) on the mean daily global insolation ranges between 0.75 (vertical tilt) and 1.43 MJ/m² (sun-tracking orientation) (or about 4.7–9.0 kW h/kW in terms of PV potential), or from 5.6% to 6.9% of the mean. Ultimately insolation and photovoltaic potential were mapped over the country at a 300 arc seconds (～10 km) resolution. The maps are available on a Natural Resources Canada Website. This is an important new tool to help Canadians gain an overall perspective of Canada’s photovoltaic potential, and allow estimation of potential photovoltaic system electricity production at any chosen location.     

Remaining useful life prediction of PEMFC based on long short-term memory recurrent neural networks

To solve the prediction problem of proton exchange membrane fuel cell (PEMFC) remaining useful life (RUL), a novel RUL prediction approach of PEMFC based on long short-term memory (LSTM) recurrent neural networks (RNN) has been developed. The method uses regular interval sampling and locally weighted scatterplot smoothing (LOESS) to realize data reconstruction and data smoothing. Not only the primary trend of the original data can be preserved, but noise and spikes can be effectively removed. The LSTM RNN is adopted to estimate the remaining life of test data. 1154-hour experimental aging analysis of PEMFC shows that the prediction accuracy of the novel method is 99.23%, the root mean square error (RMSE) and mean absolute error (MAE) is 0.003 and 0.0026 respectively. The comparison analysis shows that the prediction accuracy of the novel method is 28.46% higher than that of back propagation neural network (BPNN). Root mean square error, relative error (RE) and mean absolute error are all much smaller than that of BPNN. Therefore, the novel method can quickly and accurately forecast the residual service life of the fuel cell.     

ARMA based approaches for forecasting the tuple of wind speed and direction

Short-term forecasting of wind speed and direction is of great importance to wind turbine operation and efficient energy harvesting. In this study, the forecasting of wind speed and direction tuple is performed. Four approaches based on autoregressive moving average (ARMA) method are employed for this purpose. The first approach features the decomposition of the wind speed into lateral and longitudinal components. Each component is represented by an ARMA model, and the results are combined to obtain the wind direction and speed forecasts. The second approach employs two independent ARMA models – a traditional ARMA model for predicting wind speed and a linked ARMA model for wind direction. The third approach features vector autoregression (VAR) models to forecast the tuple of wind attributes. The fourth approach involves employing a restricted version of the VAR approach to predict the same. By employing these four approaches, the hourly mean wind attributes are forecasted 1-h ahead for two wind observation sites in North Dakota, USA. The results are compared using the mean absolute error (MAE) as a measure for forecasting quality. It is found that the component model is better at predicting the wind direction than the traditional-linked ARMA model, whereas the opposite is observed for wind speed forecasting. Utilizing VAR approaches rather than the univariate counterparts brings modest improvement in wind direction prediction but not in wind speed prediction. Between restricted and unrestricted versions of VAR models, there is little difference in terms of forecasting performance.     

Statistical post‐processing of turbulence‐resolving weather forecasts for offshore wind power forecasting

Accurate short‐term power forecasts are crucial for the reliable and efficient integration of wind energy in power systems and electricity markets. Typically, forecasts for hours to days ahead are based on the output of numerical weather prediction models, and with the advance of computing power, the spatial and temporal resolutions of these models have increased substantially. However, high‐resolution forecasts often exhibit spatial and/or temporal displacement errors, and when regarding typical average performance metrics, they often perform worse than smoother forecasts from lower‐resolution models. Recent computational advances have enabled the use of large‐eddy simulations (LESs) in the context of operational weather forecasting, yielding turbulence‐resolving weather forecasts with a spatial resolution of 100 m or finer and a temporal resolution of 30 seconds or less. This paper is a proof‐of‐concept study on the prospect of leveraging these ultra high‐resolution weather models for operational forecasting at Horns Rev I in Denmark. It is shown that temporal smoothing of the forecasts clearly improves their skill, even for the benchmark resolution forecast, although potentially valuable high‐frequency information is lost. Therefore, a statistical post‐processing approach is explored on the basis of smoothing and feature engineering from the high‐frequency signal. The results indicate that for wind farm forecasting, using information content from both the standard and LES resolution models improves the forecast accuracy, especially with a feature selection stage, compared with using the information content solely from either source.     

基于CNN-Bi-LSTM的太阳辐照度超短期预测

针对太阳辐射引起光伏出力的不确定性和波动性,进而造成大量光伏发电并网时对电网稳定性和安全的危害,提出一种新的太阳辐射超短期预测方法.该方法通过构建一维卷积神经网络,对多个关键气象变量进行数据融合和特征转换,然后构造双向长短期记忆网络预测模型,实现对未来15 min的太阳总辐照度的超短期预测.实验结果表明,所提出的预测模...     

Two estimation methods for monthly mean hourly total irradiation on tilted surfaces from monthly mean daily horizontal irradiation from solar radiation data of Ajaccio, Corsica

The sizing of a photovoltaic or a thermal solar system is generally based on monthly mean values of daily solar radiation on tilted surfaces. Many authors have demonstrated that it will be better to use monthly mean values of hourly radiation, particularly taking into account the Sun's position and to predict long-term performances of solar systems. (Liu and Jordan, 1963; Clark et al., 1984). Moreover, for most of the sites around the world, only monthly mean values of daily horizontal total irradiation are available for use in such calculations. We propose, by using well-known correlations in the literature, to estimate these monthly mean values of hourly total irradiation on tilted planes from monthly mean values of daily total horizontal irradiation, using three steps:

• — determination of monthly mean value of hourly total horizontal irradiation;

• — determination of monthly mean value of hourly diffuse horizontal irradiation;

• — determination of monthly mean value of hourly total irradiation on tilted planes.

In the first step, using the Collares Pereira and Rabl correlation, the root mean square error (RMSE) between correlated and experimental calculated data is 8%. In the second step, we used two methods: the first one utilizes the Erbs correlation and the second one is based on a local correlation which has been developed in our centre. Both of them gave identical results with an RMSE lower than 9%. We calculated monthly mean values of hourly total irradiation on three tilted planes (30°, 45° and 60°) and we compared these results with the experimental ones, obtaining a RMSE respectively of less than 10%. The method is then validated by these results.     

Residential PV capacity estimation and power disaggregation using net metering measurements

As the intermittency and uncertainty of photovoltaic (PV) power generation poses considerable challenges to the power system operation, accurate PV generation estimates are critical for the distribution operation, maintenance, and demand response program implementation because of the increasing usage of distributed PVs. Currently, most residential PVs are installed behind the meter, with only the net load available to the utilities. Therefore, a method for disaggregating the residential PV generation from the net load data is needed to enhance the grid-edge observability. In this study, an unsupervised PV capacity estimation method based on net metering data is proposed, for estimating the PV capacity in the customer’s premise based on the distribution characteristics of nocturnal and diurnal net load extremes. Then, the PV generation disaggregation method is presented. Based on the analysis of the correlation between the nocturnal and diurnal actual loads and the correlation between the PV capacity and their actual PV generation, the PV generation of customers is estimated by applying linear fitting of multiple typical solar exemplars and then disaggregating them into hourly-resolution power profiles. Finally, the anomalies of disaggregated PV power are calibrated and corrected using the estimated capacity. Experiment results on a real-world hourly dataset involving 260 customers show that the proposed PV capacity estimation method achieves good accuracy because of the advantages of robustness and low complexity. Compared with the state- of-the-art PV disaggregation algorithm, the proposed method exhibits a reduction of over 15% for the mean absolute percentage error and over 20% for the root mean square error.     

A statistical approach for sub-hourly solar radiation reconstruction

This paper proposes a computational-statistics based approach for solar radiation reconstruction at sub-hourly intervals. A dimensionless form of stochastic variable, V, which is defined as the difference between the theoretical global solar radiation in clear-sky conditions and the actual solar radiation, normalized by the clear-sky global solar radiation, is introduced and adopted in this work. The probability density function of V is calculated from historical data using a Gaussian kernel density estimator. With the developed model, the only input information required for the reconstruction procedure is the cloud condition of the sky (i.e., fair, partly cloudy, overcast, and rain/snow etc.). A case study in simulating solar radiation in Singapore is conducted to validate the accuracy of the model. The calculated results agree well with the measured data. The normalized root mean square error (NRMSE) is on average 23.4% and 7.2% for the one-minute temporal resolution and hourly integral values, respectively.     

Online 24-h solar power forecasting based on weather type classification using artificial neural network

Power forecasting is an important factor for planning the operations of photovoltaic (PV) system. This paper presents an advanced statistical method for solar power forecasting based on artificial intelligence techniques. The method requires as input past power measurements and meteorological forecasts of solar irradiance, relative humidity and temperature at the site of the photovoltaic power system. A self-organized map (SOM) is trained to classify the local weather type of 24 h ahead provided by the online meteorological services. A unique feature of the method is that following a preliminary weather type classification, the neural networks can be well trained to improve the forecast accuracy. The proposed method is suitable for operational planning of transmission system operator, i.e. forecasting horizon of 24 h ahead and for PV power system operators trading in electricity markets. Application of the forecasting method on the power production of an actual PV power system shows the validity of the method.