TMY 2011 ΔTMY 2012 ΔTMY 2013 ΔTMY 2014 ΔTMY GH
4. RESULTS AND DISCUSSIONS
4.1.1 Novel short-term forecasting method findings
The investigation on short-term forecasting using the described Hybrid method in 3.3.2 was first presented as a concept of using boundary meteorological stations in Singapore in an “alarm detection system” for approaching storms, which in turn helped reduce forecasting errors in a tropical location like Singapore (NOBRE, SEVERIANO_JR., et al., 2014). In more recent investigations on the topic, two other extreme weather phenomena were added to the analysis – “washout days” (days with extreme rain events) and also “hazy days” (where air pollution was considerably high). The work was also further expanded with the addition of data from 2014, together with the previously presented 2013 case studies in order to further strengthen the applicability and robustness of the envisioned method.
As described in 3.3.2, ten random days, among the portfolio of stormy days, were selected for the evaluation of the algorithm. The findings for the first case study analysis are summarized in Table 16. It shows results for the ten selected “sudden storm” days. The proposed Hybrid method achieved performance improvements of ~10% in absolute terms if compared with the Persistence method as a forecasting baseline. Also noticeable is the fact that ARIMA underperformed against the Persistence model for days with extreme weather conditions. This can be explained by the classical storm duration in Singapore starting in mid- or late afternoon, with dissipation near sunset or shortly after it. This allows the Persistence forecast to thrive, which is one of the reasons for this method being chosen as an algorithm of choice for the Hybrid model after a storm has reached the validation station.
An example of the better performance of the Hybrid method is shown in Figure 88 (top), with the forecasted irradiance (black dash line) following the measured irradiance (green continuous line) thanks to an earlier-detected approaching storm. The minimization of the forecast error can also be visualized in Figure 88 (bottom), where ARIMA and Persistence forecasts demonstrated their known under-performances due to their reliance on past irradiance step information.
Table 16: Sudden storm forecasting error improvement when using the novel Hybrid method (NOBRE et al., 2015 (submitted)).
Figure 88: Measured GHI at a central site in Singapore, with the forward ARIMA, Persistence and proposed Hybrid 15-min forecasts (top). The MAPE is given, showing considerable error spikes upon the arrival of the storm for both Persistence and ARIMA methods, but not for the proposed Hybrid model (bottom) (NOBRE et al., 2015 (submitted)).
Results for the second case study addressing ten random washout events are presented in Table 17, including the total daily irradiation, as well as the nRMSEs obtained by using the Persistence-only and ARIMA- only forecast methods. In such example data, the Persistence forecast vastly outperformed ARIMA, which would be expected for a situation of total cloud cover condition (very low kt values), which suits Persistence. The error avoidance column aims at highlighting detrimental errors for non-adaptive short-term forecasting models, which rely on running a single algorithm throughout all daily and yearly conditions.
Table 17: Case study on washout conditions in Singapore and error avoidance by selection of the Persistence method over ARIMA (NOBRE et al., 2015 (submitted)).
Finally, the third set of case studies is aimed at simplifying short- term irradiance forecasting during strong haze episodes, acting similarly to the washout events. The PSI threshold criterion of 100 was easily crossed in 2013, when Singapore experienced a notorious haze season, reaching unhealthy air quality levels for several days. In 2014, the number of occurrences of haze days was lower.
Table 18 shows the results of Persistence-only versus ARIMA- only forecasts. The error avoidance metric did not indicate much lower values for the hazy, as they were for the washout days (with the use of Persistence), but indicated nevertheless that future short-term forecasting methods may as well use the more simplistic Persistence forecast during such air pollution events. In any case, it is important to remember that haze comes associated with loss of irradiation at ground level. Thus, a limitation in the maximum output of PV systems must be accounted for,
also when modeling its grid integration impact due to the energy generation shortfall expected.
Table 18: Case study on hazy days in Singapore and error avoidance by selection of the Persistence method over ARIMA (NOBRE et al., 2015 (submitted)).
The novelty of the proposed method lies on the prediction of massive drops in irradiance (i.e. consequently power output of PV installations) caused by approaching storms approximately 15-30 minutes in advance – which is typically sufficient to ramp-up conventional generation capacities (in Singapore, gas-fired turbines). Furthermore, the sensor network allows the identification of washout conditions before sunrise, allowing the grid operator to also react in advance and compensate for the missing solar power in such days.
Most cloud motion tends to come from the West parts of Singapore (thus coming from Sumatra in Indonesia, as per example animation of the Sumatra squall in Figure 11). The caveat: it is expected that there is a tendency for better forecasts for the Eastern parts of the island as most cloud activity for Singapore initiates from the West.