Estimating Real Time GB Aggregated Solar PV ...

2 downloads 64 Views 1MB Size Report
energy sources, Solar energy, Solar power generation. I. INTRODUCTION ... sites, NG's daytime demand forecast error has increased steadily in line with solar ...
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT)
REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < half hourly input data would be in excess of 50% using ground mounted irradiance data from the UK Met Office [5] (see Fig. 13 of Appendix). Furthermore, real-time weather data of sufficient resolution may be expensive or unavailable. Here we present an alternative statistical approach to modelling the generation of an ensemble of PV generators, whereby we use real-time generation data from a sample of real PV systems. Some examples of such an approach can be found in the literature, with authors having considered both Kriging [6] and Neural Network [7] approaches, but in general, the availability of real-time PV generation data has been insufficient to achieve a workable solution on a national scale at half-hourly resolution. In the literature, there are some examples of authors studying similar approaches to estimating aggregate PV generation [3] [8], but to the best of our knowledge the methodology documented here is original and the resulting service, PV_Live, is unique in the UK. The proposed methodology and PV_Live service is valuable to the decision making of TSOs and DNOs as well as energy generators and traders. For TSOs, this algorithm provides visibility of the outturn of distributed solar PV connected to the electricity network, enabling them to improve their PV outturn forecasts and as such make more informed decisions regarding the trading of electricity and their ‘operating reserve’. For DNOs, the algorithm can provide insight into which sections of their distribution network are likely to require attention with regards to PV deployment, be it reinforcing the grid or suspension of PV deployment in that particular area. For energy traders, knowledge of the real outturn is crucial to developing and validating accurate PV generation forecasts, which in turn help to trade electricity more effectively.

2

in reality a “real-time” data feed is not feasible due to the source of the data and the time taken to download. The 5 minutely generation data from these sites is downloaded every 15 minutes from a web service (Application Programming Interface, API) and aggregated to half hourly, since this is the standard settlement period of interest to the renewables forecasting team at NG. The time needed to download and preprocess the data means that for a particular half hour period the data is available in a local database within 12 minutes after the end of the half hour. Considering the potential for network latency, we allow an additional 3 minutes after the end of each half hour period before initiating the modelling process, bringing the total latency to 15 minutes.

II. METHOD The difficulty in modelling PV generation stems from the geospatial and temporal variability of weather combined with the variability in performance of PV systems as a function of solar resource, weather, system configuration, technology, and installation quality. By using real world PV generation as an input to our model and ensuring a representative sample, we are able to isolate many of the latter effects i.e. working under the assumption that the performance of our sample is representative of GB PV performance as a whole. Our approach uses the available sample of real-time PV generation to estimate a representative performance factor for the entire GB fleet, which can then be scaled to the total effective capacity of solar PV. A. Real time data One of the challenges associated with real time data provision is acquiring real-time input data. By leveraging donors from the Microgen Database [9] and sourcing new commercial donors, we have acquired near live data from 627 sites across the UK. This is then processed in real time. The data consists of 5 minutely average power output readings which are converted to half hourly integrated energy readings during a pre-processing stage. The term ”real-time” is somewhat of a misnomer here, since

Fig. 1. Map of all reported PV systems as of 2016-07-29 according to NG.

B. Representative sampling The distribution of solar PV systems in the UK is nonuniform and the installed capacity shows a significant trend toward more deployment being in the South of the country, see Fig. 1. In order to estimate the representative performance of all PV systems in the GB fleet, we select an optimal subsample of sites from our available real-time sites in order to maximise the geographical representativeness with respect to all reported sites (Fig. 2). To achieve this, we define a two dimensional Chisquared metric to measure geographical representativeness in terms of effective capacity density. Fig. 3 demonstrates how the 2D Chi-squared metric is calculated. The width and height of the cells in Fig. 3 can be tuned to

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT)
REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < The unreported capacity is categorised as “Revision”, “Uplift” and “Unregistered”:  Revision - Capacity that has already been commissioned but is missing from the latest reports due to delay in submitting and processing applications.  Uplift - Capacity that has been commissioned since the latest report(s) were compiled. Most reporting mechanisms publish reports quarterly.  Unregistered - Capacity that has been commissioned but is not registered with any of the reporting mechanisms considered. E. Uncertainties Understanding and quantifying the uncertainties on the PV_Live results is crucial if they are to be used effectively by demand forecasters at NG and other industry stakeholders. Understandably, it is difficult to derive and validate the uncertainty on a national aggregate of PV generation when there is no directly measured equivalent data to compare to. Instead we rely upon our ability to estimate the uncertainty at each stage in the PV_Live calculation and propagate those errors through to the final result. Here we will consider only the dominant uncertainties in the results, since a more detailed exploration of the uncertainties will be the subject of future work. 1) Uncertainty in installed capacity In Fig. 4, we highlight the components of the installed capacity derivation that are subject to uncertainty, which we will now attempt to quantify. We have made use of historic Feed In Tariff (FIT) reports published by OFGEM [12] in order to estimate the lag in this reporting mechanism and to make assumptions about the magnitude and allocation of revision capacity. The corresponding uncertainty on revision capacity is estimated as [-5%, +10%]. Using linear extrapolation of monthly aggregated installed capacity, we estimate the uplift capacity in July and August 2016 to be 265 and 308 kWp respectively, with an estimated uncertainty of [-10%, +30%]. The uncertainty due to unregistered capacity has been conservatively estimated as +1%. Combining these uncertainties we estimate the overall uncertainty on the cumulative installed capacity estimate as of 31st August 2016 to be [-2%, +5%], i.e. 11.4 +0.5 −0.2 GWp.

4

2) Uncertainty on representative yield We estimate uncertainty in the representative yield calculation in several ways. Firstly, we can observe the standard error on the mean (3) of the optimised sample yield. Similarly, we can observe the population standard deviation of optimised sample yield. These metrics give some indication of the realtime variability in PV generation, but a more useful description of the uncertainty for NG is a confidence interval of particular statistical significance. We make use of the 90% Bayesian confidence interval for the mean yield of the sample [14]. By normalising each of these metrics to the mean yield, each uncertainty metric can be easily combined with the uncertainty on the installed capacity. In the case of standard error and standard deviation, the errors should be added in quadrature.

𝑆𝐸𝑥̅ =

𝜎 √𝑛

(3)

Where:

𝑆𝐸𝑥̅ 𝜎 𝑛

Standard Error on the mean Sample standard deviation Sample size

F. Real time provision of results In order to provide PV_Live results in real-time to NG and other stakeholders, we upload results to a dedicated server from which results are accessible over the world wide web via a convenient API. We also publish the results on the Sheffield Solar website [15], and several third-party websites, apps and services. Details of the public API and some third party services are published on the Sheffield Solar website.

Fig. 6. A flowchart of the process undertaken each half hour in order to calculate and provision the PV_Live results.

III. RESULTS AND VALIDATION Validation of a national aggregate of PV generation is difficult due lack of exhaustive direct metering of distributed PV generation. We have considered two routes to validating the real-time results provided by PV_Live.

Fig. 5. Growth in installed capacity of solar PV in the UK since 1st January 2015. Red shaded area shows the uncertainty estimate. Installed capacity has more than doubled since the start of 2015.

A. Demand Forecast Error validation NG forecast nationally aggregated electricity demand by modelling so-called ‘cardinal points’, which represent turning points in the electricity demand profile within particular time intervals. The demand between cardinal points is then estimated

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < using well established interpolation techniques [16]. Naturally, the demand forecast error (DFE) is examined per cardinal point, and in the case of PV generation, only cardinal points that occur during daylight hours are considered. One route to validation of PV_Live results is to study the DFE before and after the introduction of PV_Live results to the demand forecast models. In doing so, we can identify from NG’s perspective whether the results offer improved accuracy over existing methodologies, and also ensure that successive iterations of the PV_Live algorithm outperform the last. In Fig. 7 we present the DFE of the most recent PV_Live version (0.1.5b) compared with the benchmark DFE for the existing, pre-PV_Live model. The results show a reduction in DFE of up to 130 MW since the introduction of PV_Live. The demand forecast models developed by NG are complex, with many control variables, and as such it would be inaccurate to claim that this reduction is solely due to the introduction of PV_Live results.

5

to this project is estimated to be worth £1m - 10m per annum. To reinforce this point, Fig. 8 shows how the midday demand varies as a function of PV generation, broken down by year and time zone. Since distributed PV generation effectively reduces electricity demand (INDO) as observed by the TSO, one would expect the gradient of the line of best fit for each graph to be −1. In reality, the demand during 2013/2014 is predominantly influenced by other factors, hence the poor fits in 2013/2014, especially during GMT. As UK PV deployment becomes more substantial in 2015/2016, we can see that the PV generation begins to have significant impact on the observed demand, with the gradient tending towards −1. The implication for the TSO is that accurate indirect metering of PV generation is of vital importance if they are to accurately monitor and forecast the metered demand outturn.

Fig. 8. Initial Network Demand Outturn against PV generation during the interval 11:30 – 12:30 (GMT), by year and time zone. Colouring shows nonparametric density of points and the gradient in each subplot refers to the gradient of the black line of best fit. Only non-bank-holiday weekdays are included due to the disparity in demand between weekdays and weekends/bank holidays.

Fig. 7. Root Mean Square Error (RMSE) of NG’s transmission demand forecast before and after the introduction of PV_Live.

B. Cost Savings To convert DFE into the cost saving for the consumer is nontrivial, since the relationship between demand forecast uncertainty and cost to the consumer is complex. NG compensates for demand forecast uncertainty by using the Short Term Operating Reserve (STOR) mechanism to reserve generation capacity or demand reduction that can be activated with short notice. The STOR market alone cost £49 million in 2015/2016 [17]. A long term reduction in demand forecast can theoretically lead to a reduction in the operational expenditure of the TSO on short term reserve, since in theory more capacity may be reserved in advance. In practice things are not so straightforward and the main cost benefit of PV_Live lies elsewhere. In 2014, NG estimated that 400 MW of demand forecast error, as was typical prior to 2015 when installed capacity of PV was less significant, equates to roughly £100 million per annum in operational reserve. Since the growth in deployment of PV is continuing despite cuts to Government financial incentives, this trend suggests NG’s DFE would continue to rise, which would surely lead to increased operational reserve costs. In undertaking this work we pave the way for NG to offset this future increase by providing improved monitoring of PV generation which in turn enables improved demand forecasting. The offset of future operational costs due

C. Further validation A second approach to validation of the real-time results involves retrospectively re-running the calculations with an increased available sample size. This assumes that uncertainty in the representative yield calculation will decrease as more systems are sampled. Validations of this kind will be the subject of future work in this project, and we are continually seeking to expand the database of historic PV generation data for this purpose. This type of validation can help to steer development of the algorithms and ensure that successive versions of PV_Live provide improved accuracy. Similarly, we will seek to further develop our understanding of uncertainty on the installed capacity data by studying subsequent reports to assess whether our uncertainty estimates were reasonable, whilst also exploring new sources of information regarding PV deployment. D. GB PV Statistics As well as providing crucial visibility to NG’s demand forecasters, PV_Live also provides an opportunity to characterise aggregate PV generation in Great Britain, providing insight into the role of PV in the GB energy mix. Typical profiles for Nationally aggregated PV generation can be seen in Fig. 9.

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT)
REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < ACKNOWLEDGMENT

[14] T. E. Oliphant, “A Bayesian perspective on estimating mean,

We wish to thank Jack Barber for his contribution to the early development of the methodology discussed. Sheffield Solar wish to thank Samuel Chase for administering the PV data collection and pre-processing.

[15]

REFERENCES

[16]

[1]

[2] [3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

National Grid plc, “PV Monitoring Phase 2 Project Registration Document,” 16 09 215. [Online]. Available: http://www.smarternetworks.org/NIA_PEA_PDF/NIA_NGET 0170_4703.pdf. [Accessed 18 11 2016]. A. P. Dobos, “PVWatts Version 5 Manual,” NREL, 2014. D. Lingfors and J. Widén, “Development and validation of a wide-area model of hourly aggregate solar power generation,” Energy, vol. 102, pp. 559-566, 2016. G. Colantuono, A. Everard, L. M. Hall and A. R. Buckley, “Monitoring nationwide ensembles of PV generators: Limitations and uncertainties. The case of the UK,” Solar Energy, vol. 108, pp. 252-263, 2014. Met Office, NCAS British Atmospheric Data Centre, “Met Office Integrated Data Archive System (MIDAS) Land and Marine Surface Stations Data (1853-current),” 2012. [Online]. Available: http://catalogue.ceda.ac.uk/uuid/220a65615218d5c9cc9e4785a 3234bd0. [Accessed 11 09 2015]. Y. M. Saint-Drenan, S. Bofinger, B. Ernst, T. Landgraf and K. Rohrig, “Regional nowcasting of the solar power production with PV-plant measurements and satellite images,” in 30th ISES Biennial Solar World Congress 2011 Vol 2, Kassel, Germany, 2011. A. G. R. Vaz, B. Elsinga, W. G. J. H. M. van Sark and M. C. Brito, “An artificial neural network to assess the impact of neighbouring photovoltaic systems in power forecasting in Utrecht, the Netherlands,” Renewable Energy, vol. 85, pp. 631-641, 2016. Y. M. Saint-Drenan, G. H. Good, M. Braun and T. Freisinger, “Analysis of the uncertainty in the estimates of regional PV power generation evaluated with the upscaling method,” Solar Energy, vol. 135, pp. 536-550, 2016. Sheffield Solar, “Microgen Database,” Sheffield Solar University of Sheffield, [Online]. Available: http://www.microgen-database.org.uk/. D. C. Jordan and S. R. Kurtz, “Photovoltaic Degradation Rates — An Analytical Review,” Progress in Photovoltaics: Research and Applications, vol. 21, no. 1, pp. 12-29, 2013. J. Taylor, J. Leloux, L. M. H. Hall, A. Everard, J. Briggs and A. R. Buckley, “Performance of Distributed PV in the UK: A Statistical Analysis of Over 7000 Systems,” in 31st European Photovoltaic Solar Energy Conference and Exhibition, Hamburg, 2015. UK Office of Gas and Electricity Markets (OFGEM), “Installation reports,” [Online]. Available: https://www.ofgem.gov.uk/environmentalprogrammes/fit/contacts-guidance-and-resources/publicreports-and-data-fit/installation-reports. [Accessed 20 11 2016]. Department for Business, Energy & Industrial Strategy, “Renewable energy planning database monthly extract,” [Online]. Available: https://www.gov.uk/government/publications/renewableenergy-planning-database-monthly-extract. [Accessed 20 11 2016].

7

[17]

[18]

variance, and standard-deviation from data,” All Faculty Publications Paper 278, 2016. University of Sheffield, “PV_Live,” Sheffield Solar, 1 11 2015. [Online]. Available: https://www.solar.sheffield.ac.uk/pvlive/. [Accessed 23 11 2016]. J. W. Taylor and S. Majithia, “Using combined forecasts with changing weights for electricity demand profiling,” Journal of the Operational Research Society, vol. 51, no. 1, pp. 72-82, 2000. National Grid plc, “Services Reports MBSS_MARCH_2016(1),” 11 05 2016. [Online]. Available: http://www2.nationalgrid.com/UK/Industryinformation/Electricity-transmission-operational-data/Reportexplorer/Services-Reports/. [Accessed 20 11 2016]. Western Power Distribution, “Report on the use of proxy PV FiT meters to reflect local area,” 05 2013. [Online]. Available: https://www.westernpowerinnovation.co.uk/Documentlibrary/2013/PvFITProxy-final-submitted-v1-May-13.aspx. [Accessed 20 11 2016].