A Novel Method for Predicting the Power Output of Distributed Renewable Energy Resources Athanasios Aris Panagopoulos1 Supervisor: Georgios Chalkiadakis1 Technical University of Crete, Greece A thesis submitted to the Department of Electronic and Computer Engineering in partial fulfillment of the requirements for the degree of Diploma in Engineering 1 Electronic and Computer Engineering, Technical University of Crete, Greece; emails: {apanagopoulos, gchalkiadakis} @isc.tuc.gr 1 of 30
Motivation Renewable energy sources need to get integrated into the electricity grid: Inherently Intermittent Potentially Distributed Smart Grid Technologies are the key for: The successful integration of the numerous distributed energy resources Decision-making regarding energy production and/or consumption 2 of 30
Virtual Power Plants (VPPs) AI and MAS research has been increasingly preoccupying itself with building intelligent systems for the Smart Grid Virtual Power Plants (VPPs) Coalitions of energy producers, consumers and/or 'prosumers' e.g. wind turbines, solar panels, electric vehicles batteries 3 of 30
Virtual Power Plants (VPPs) AI and MAS research has been increasingly preoccupying itself with building intelligent systems for the Smart Grid. Virtual Power Plants (VPPs) Coalitions of energy producers, consumers and/or 'prosumers' e.g. wind turbines, solar panels, electric vehicles batteries Equipping VPPs with an algorithmic framework and a web-based tool for dependable power output prediction of Photovoltaic Systems (PVSs) and Wind Turbines Generators (WTGs) across the Mediterranean Belt Our methods use free-to-all meteorological data 4 of 30
PVS Power Output Prediction Forecasting PV systems output can be linked to the task of forecasting solar irradiance estimates. Drawbacks of existing approximation methods: They rely on expensive meteorological forecasts. Many such methods produce clear sky prediction models only Usually no strict approximation performance guarantees 5 of 30
PVS Power Output Prediction Forecasting PV systems output can be linked to the task of forecasting solar irradiance estimates. Drawbacks of existing approximation methods: They rely on expensive meteorological forecasts. Many such methods produce clear sky prediction models only Usually no strict approximation performance guarantees They are made up of components that have been evaluated only in isolation Their performance has been evaluated only in a narrow geographic region Examples: SVMs, MLP networks etc 6 of 30
Overview of Main Contributions Novel non-linear approximation methods for solar irradiance falling on a surface, given cloud coverage A generic PVS power output estimation model combining our solar irradiance model with existing models calculating various PV systems losses Cheap methods: only require weather data readily available to all for free, via meteo websites Methods applicable to a wide region Evaluation based on real data coming from across the Mediterranean belt (Med-Belt) Error propagation procedure to estimate our method s total error for the entire Med-Belt RENES: a web-based, interactive DER output estimation tool incorporates our PVS power output estimation methods also produces WTG power output estimates 7 of 30
Overview of Main Contributions First work to use a generic and low-cost methodology incorporating solar irradiance estimation and free-to-all weather data Evaluated in a wide region RENES: a web-based, interactive DER output estimation tool: Incorporates our PVS power output estimation methods Also produces WTG power output estimates RENES: a convenient user-interactive tool for: simulations and experiments comes complete with an API and XML responses VPPs operating A paper based on this work, entitled Predicting the Power Output of Distributed Renewable Energy Resources within a Broad Geographical Region and co-authored by ArisAthanasios Panagopoulos, Dr. Georgios Chalkiadakis and Dr. Eftichios Koutroulis, was awarded the best student paper award in the Prestigious Applications of Intelligent Systems (PAIS) track of the 2012 European Conference on Artificial Intelligence (ECAI 2012) 8 of 30
A PVS Power Output Estimation Model The method for predicting the energy output of PV systems consists of the estimation steps: I. Developing a solar irradiance model to predict the incident radiation,, on the PV module II. Estimating the amount of incident radiation actually absorbed by the PV module, III. Predicting the module s operating temperature, IV. Calculating the PV module s maximum power output, V. Predicting the PV system s actual power output, 9 of 30
A PVS Power Output Estimation Model The method for predicting the energy output of PV systems consists of the estimation steps/submodels: 10 of 30
An All-Sky Solar Irradiance Model stands for the total incident radiation on an arbitrarily oriented surface given a cloud coverage level N. It consists of three components: Beam Sky-diffuse Ground reflected. 11 of 30
An All-Sky Solar Irradiance Model++, \.. For β=0 12 of 30
Estimating the Cloud Transmittance Coefficients For we need to estimate the cloud transmittance coefficients Note that: There is no direct way to calculate However and and measurements are relatively commonplace I. Develop a Cloud cover Radiation Model (CRM), to estimate II. Decompose the estimated known Diffuse Ratio Model (DRM) back to and, employing a 13 of 30
Non-Linear Equation Models (CRM) They attempt to approximate the and solar elevation) ratio (as it is independent of the season Coefficients determined through least-squares fitting 14 of 30
Informed Non-Linear Equation Models (CRM) Air transparency depends on dew point temperature The difference between season/time: and is expected to be less dependent on location and Incorporating Coefficients determined through least-squares fitting in our model: 15 of 30
An MLP Network We also trained a MLP neural network with one hidden layer The network computes the quantity given: The level of cloud coverage, N The estimated The environmental temperature, The relative humidity, RH quantity (in components) 16 of 30
Our Cloud Cover Radiation Model (CRM) The nine (9) CRM approaches are: Four (4) non-linear equation models Four (4) informed non-linear equation models, trained on top of the simple non-linear equation models An MLP network Trained and evaluated with the purpose of adopting one for our CRM in our region of interest 17 of 30
Incorporating Real Data Meteorological data drawn from the Weather Underground database for 9 regions in the MedBelt, and 1 region in Northern Europe: sky condition (qualitative observations) solar radiation (i.e., ) ambient temperature ( ) relative humidity (%) I. At least one year worth of observation data during 2009-2012 was collected in each city II. Quality control tests were performed III. Reduction of the larger datasets by progressively retaining every second observation IV. All Med-Belt sets were collated 18 of 30
The Final Dataset From this we derive with training, testing and validation set 19 of 30
Least-Squares Fitting and MLP Training Least-squares fitting of the non-linear curves Choice of unique mid-point quantitative values to characterize each cloud coverage level Computation of the sample mean of the corresponding for each of these values of N Least square fitting over the pairs 20 of 30
Least-Squares Fitting and MLP Training Least-squares fitting of the informed non-linear curves Choice of unique mid-point quantitative values to characterize each cloud coverage level Computation of the sample mean of the corresponding those values of N. Least square fitting over the pairs for each of. 21 of 30
Least-Squares Fitting and MLP Training Training the MLP network MLP comprising of 4 nodes in the hidden layer was found to present the best network architecture. Normalized values in the range of [-1,1] for the quantities at the input nodes The MLP training used the back propagation learning algorithm with the batch method and uniform learning Overfitting is avoided via the early stopping neural network training technique 22 of 30
Evaluating Cloud Cover Radiation Model (CRM) Comparison outside the Med-Belt ANOVA Tests Local Training and Evaluation 23 of 30
Local Cloud Cover Radiation Model (CRM) Example results: 24 of 30
Error Propagation Methodology 25 of 30
Error Propagation Methodology 26 of 30
Error Propagation Methodology Final power output prediction performance guarantees: South-facing, 45 slope angle orientation: worst-case bound for rmae in the order of 40% No known comparable generic prediction methodology for the Med-Belt Our method s irradiance forecasting error is comparable to or lower than that of several other more expensive methods evaluated in southern Spain 27 of 30
WTG Power Output Estimation Wind speed forecasts are commonplace WTGs power output depends on the so-called power curve Inside the range of: Cut-in wind speed limit Cut-out wind speed limit 28 of 30
RENES: A Web-Based DER Output Estimation Tool http://www.intelligence.tuc.gr/renes WTG and PVS power output prediction User clickable map Automatically populated parameters' values API with XML responses 29 of 30
Conclusions and Future Work This work is the first to provide low-cost power prediction estimates via a method applied to a wide region, incorporating solar irradiance forecasts in the process We implemented a web-based, interactive DER power output estimation tool (RENES) Method and tool are extensible Can incorporate any other intermediate-step techniques deemed appropriate for particular sub-regions RENES is a convenient user-interactive tool for simulations and experiments, and can be of use to VPPs / wider public We plan to employ RENES in VPPs-related simulations/ experiments and also to get readings for simulations related to optimal sun-tracking Thank you! Any questions? Department of Electronic and Computer Engineering, Technical University of Crete, Greece 30 of 30