Short-term PV power forecasting in India: recent developments and policy analysis

For a power system in which power generation from RE has priority feed-in, estimating the residual or net load forecast is important for the TSO. In order to generate net load forecast, control area-level aggregate generation forecast is necessary. Residual or net load forecast can be used by the TSO to schedule the dispatchable generators in order to supply the net load. It is also useful for ascertaining that sufficient ramping capability is available to compensate for the ramps in net load. The uncertainty information associated with net load can be used as an input into reserve requirement estimation [96]. Nodal (pooling station level) forecasts are useful to the TSOs for both intra- and inter-control area congestion forecast and management [57]. Power markets are also affected by network bottlenecks and hence can utilize nodal forecast information [7] for cases where power markets set nodal energy prices [63, 87]. Individual power plant-level generation forecast and its associated uncertainty can be of use to the TSO, in cases where gross pool mechanism [47] is followed or mandated by regulation, for scheduling the generators.

Steep ramps in solar PV plant power output during morning and evening times can lead to increased cycling of the dispatchable generation units, and put more stress on the power system [89]. Flexibility is necessary in order to deal with power ramps, and the requirement of flexibility also needs to be inferred from ramp forecasts or ramp alerts [86]. Many authors have described a trinity of metrics, namely, ramp rate, ramp magnitude and energy, for quantifying the ramp in power from VRE units [123]⁠.

Plant-level PV generation forecast information is useful to the plant operator for scheduling the expected generation with the grid operator. For solar PV plants participating in day-ahead power market, day-ahead generation forecast is a vital source of information for placing bids [129]. Intra-day or imbalance markets provide opportunity to the solar PV plant operators or their traders to revise their bids based on updated and more accurate intra-day forecasts [64]. Apart from single-valued deterministic forecast, forecast uncertainty information can be utilized to optimize the bids for power in the day-ahead and intra-day markets. This is especially the case when differential penalties are applied for over and under generation [23, 129].

VRE units, such as solar PV, are increasingly being expected to provide ancillary services to the grid operator, similar to conventional generating units. The ability of solar PV plants to provide ancillary services has already been documented in literature [29, 80] and demonstrated in field test [85]. Wind turbines are already being used for providing negative tertiary reserve. The role of accurate generation forecast and its uncertainty information is crucial here [53]⁠.

VPP is an aggregator of generation units, and in certain cases load entities as well. According to [107]⁠, “a VPP aggregates the capacity of many diverse DERs, creates a single operating profile from the composites of the parameters characterizing each DER and incorporates the impact of the network on aggregate DER output. A VPP is a flexible representation of a portfolio of DER that can be used to make contracts in the whole-sale market and to offer services to the system operator”. [105] defined VPP as “a portfolio of DERs, which are connected by a control system based on information and communication technology. The VPP acts as a single visible entity in the power system, is always grid-tied and can be either static or dynamic”. Individual and aggregate forecast information is necessary for VPPs to provide optimal bids in power markets, as well as for providing ancillary services to the grid operator [56, 72, 109]⁠.

Historically, distribution system has played a passive role, functioning only as the interface for the transfer of power from transmission network to end-consumers. However, with the addition of numerous distributed generators such as grid connected rooftop PV to the medium- and low-voltage distribution grid, passive operation is no longer possible. This can in turn pose barriers to the secure and economic operation of the distribution grid [46]. There is an ever-growing need for an intermediate entity between the TSO and the DERs, due to the limited visibility and control of TSO over these resources. Two primary roles that this entity (called DSO) is expected to fulfill are: (i) operator of the distribution grid and (ii) Facilitator of market access to DERs [5]⁠. One of the main challenges is to develop suitable tools and systems, which would allow DSOs to make informed decisions based on data from various sources [60], including generation forecasts, in order to fulfil their various roles. Voltage and current limitations also act as obstacles to the distribution grid’s PV hosting capacity and is the primary cause of overvoltage during high solar PV generation at noon-time [120]. As can be seen in the relevant literature [1, 15, 67, 118]⁠, several authors are of the opinion that DSOs need to start scheduling their system (including demand, generation and storage), and also manage voltage and power flow violations with the help of available resources like OLTCs, advanced inverter control techniques, capacitor banks, etc. However, for all these cases, generation forecast is the basic building block based on which the operational planning is done.

Solar PV power forecasting comprises of several steps which convert NWP output, satellite images, real-time online measured data or sky images into a forecast product, by utilizing various models. The initial set of steps for deriving the forecasted surface GHI depends on the source of data and the required spatio-temporal resolution. Further steps are necessary for producing the final AC power forecast as shown in Fig. 1. The measured data of GHI and AC power output are often used for site-adaptation in order to enhance the accuracy of the prediction in real operating systems. Brief descriptions of the different solar irradiance forecasting, forecast combination, and up-scaling techniques are provided in the following subsections.

In order to better handle the intermittency of VRE generation, while allowing its economic and secure integration, regulatory reforms have been introduced from time to time in different power systems. In Germany, the EEG act accords priority dispatch to renewable generation and makes the TSO responsible for selling power from price-taking VRE generators at the energy exchange, who are then paid under the FIT scheme. Aggregate control area-level VRE generation forecasts are utilized by TSOs for grid security assessment and aggregate VRE generation bidding at the exchange. The imbalance costs of such VRE generators are borne by the TSOs. “Price-making” VRE generators on the other hand participate in the energy exchange like any other entity and require site-specific generation forecasts. They are themselves responsible for generation forecasting and settlement of imbalances [36]. The Spanish Royal Decree RD 436/2004 introduced mandatory generation forecast for special regime RES-E generators or VRE generators, of installed capacity equal to or greater than 10 MW. This limit was further reduced by subsequent regulations. At the same time, the Spanish TSO REE also produces its own aggregate generation forecast for the Spanish electricity grid [62]. Furthermore, all VRE generators of capacity greater than 10 MW need to be connected to a centralized control centre as a necessary requirement for obtaining FIT or market premium. RD 1565/2010 has also included clusters of VRE generators connected to the same evacuation line of combined capacity greater than or equal to 10 MW within the ambit of the central control centre [32]. FERC Order 764 enabled VRE generators in USA to better manage their imbalances by introducing sub-hourly scheduling. It also allows TSOs to access sub-hourly forecast-relevant data of VRE generators for the purpose of producing power production forecasts [18, 95]. The AEMO has installed a solar power forecasting system called ASEFS under a project funded by ARENA. This was deemed essential for assisting with the operation of accurate supply and demand forecast models to increase commercial viability and ensure grid stability, in view of the large penetration of solar PV into the grid [6]⁠. The need for forecasting and scheduling of VRE generators in the Indian context was first emphasized by CERC in 2010 through the IEGC [17]⁠, and then subsequently by other SERCs and FOR, who introduced similar regulations [4, 49, 51, 66]⁠. The concept of QCA as an aggregator of VRE units connected to a pooling station as well as aggregator of multiple such pooling stations has also been introduced by FOR model regulations and other SERC regulations. A detailed analysis and inter-comparison of various aspects of the different regulations related to forecasting are included in Table 1. Recently, FOR has come up with a proposal of introducing a 5-min resolution scheduling interval in the Indian power system for higher granularity and improved ramp management [48]⁠.

Many of the SERCs have made it mandatory for all solar PV plants (with a capacity of 1/5/10 MW or above) to provide generation forecasts on a day-ahead basis to the SLDC. These regulations also allow a fixed number of intra-day revision or updates based on the latest forecasts. In particular, the requirements here are for day-ahead site-specific forecast, which can be obtained from NWP sources, and intra-day site-specific forecast, which can be generated by applying CMV to satellite images. Persistence or advanced machine learning based methods can also be used for very short-term forecast. Sky image-based methods are still at a research stage, and moreover, the use of sky imaging system for solar forecast may not be currently very useful for the Indian power system, as the minimum update interval in most of the states is 1.5 h, while the sky imaging technique can only predict for 20 to 30 minutes in the future. As per the regulations [4, 49, 66], these site-specific forecasts, different from the forecasts generated by REMCs, are provided to the SLDC for setting the schedule and consequently for determining the deviation charges.

The proposed REMCs are tasked with providing the SLDC with aggregated variable RE forecasts, which would be beneficial for the overall management of the electricity grid. The requirement here is of spatially aggregated forecast at the pooling station, state and even regional level. Such forecasts can be obtained by applying up-scaling techniques to a representative set of site-specific forecasts. State- or regional-level forecasts can find application in residual load forecasting, which would then enable state or regional LDCs to schedule the dispatchable generators for supplying the residual load. The ramps in residual load, due to the diurnal pattern of solar PV generation, can also give information on the ramping capability required from the dispatchable generators. It has been shown that VRE intermittency primarily affects the need for tertiary reserve (RRAS as per Indian power system) and not primary reserve (FGMO) [28, 70]. Thus in future, the uncertainty information associated with residual load forecast can be utilized for setting the RRAS requirement. On the other hand, nodal or pooling station-level forecasts can find application in congestion forecast and management, particularly for inter-state and inter-regional tie-lines. Load flow calculations, using nodal-level load and generation forecasts, can be employed to determine the ATC of critical transmission segments. Uncertainty information associated with nodal-level generation forecast, can also be used in future for congestion management and scheduling, as has been shown by several authors [76, 71, 59]. QCA, which has been introduced by many of the Indian states, can utilize nodal or pooling station-level aggregate forecasts for providing generation schedules to the SLDC.

According to the CEA, a distributed generator is a generating station that feeds electricity at a voltage of 33 kV or below into the electricity grid [16]. Many SERCs have already brought out draft regulations or final regulations on rooftop solar PV systems and net metering [116‐121]. Karnataka so far allows the highest penetration of solar PV up to 80% capacity of the distribution transformer level. Increase in penetration of roof-top solar PV alters the shape of the load duck curve, during the day time in particular, and thus affects the load demand scheduled by the DISCOM. To tackle this, DISCOM control centres can utilize the aggregate forecast of roof-top and distributed PV systems connected to the distribution grid. However, module information and dynamic generation data of all the solar PV units may not be available to the DISCOM directly, due to the sheer number and the spatial distribution of the units. Different up-scaling techniques can be used in such situations to obtain the aggregate output, as shown in [68, 109‐111] . Apart from this, large number of distributed solar PV plants connected to the HT and LT levels (11 and 0.4 kV) can cause spike in voltage along the feeder line, at the coupling points with the grid. In the distribution grid, active power generated by solar PV influences the voltage at the coupling point with the feeder. However, transformer with OLTC is typically only available till EHT voltage levels (33 or 66 kV) [55]. Even then, very frequent switching of the OLTC leads to wear and tear, which is also undesirable. Several methods are available in literature which show how to utilize nodal aggregated forecasts for minimizing the switching operation, while at the same time maintaining voltage limits on the different feeders [15, 114]. Some works have also proposed active power curtailment strategy for voltage regulation by utilizing solar PV generation forecasts.

Many of the applications discussed here are already used in operation internationally, and therefore in theory can also be implemented in the Indian system. Based on Table 2, a summary of the possible applications of generation forecast suggested in this section is presented along with the forecasting methods employed, its spatio-temporal resolution, and the intended end-user. However, the details for removing barriers to their implementation need further detailed investigation.

A comparative analysis of several NWP models [90] has shown that the ECMWF NWP model provides rather accurate forecast with the least uncertainty. Again, the global NWP model from ECMWF has the finest grid resolution among all the existing global NWP models. Hence for this work, the NWP model solar and cloud parameter data sets with a resolution of 0.125° × 0.125° were obtained from ECMWF for the area of the state Rajasthan. The temporal resolution of data points obtained was 3 h with a forecast horizon of 72 h.

For the investigation and the evaluation of irradiance forecasts, solar irradiance data sets were obtained from 10 SRRA stations in Rajasthan measured during the period 1 January 2014 to 31 December 2014. The distribution of these stations in Rajasthan is shown in Fig. 9.

The forecasts are to be evaluated for point locations or single sites and for an ensemble of sites with SRRA stations in Rajasthan. Irradiance values at each time step are isolated and normalized to the time interval to obtain NWP model output GHI values. The temporal resolution of the calculated NWP model GHI is 3 h.

After obtaining ECMWF GHI forecasts of 3-h temporal resolution at each grid point, averaging of forecasted GHI values over a 0.25° × 0.25° grid, i.e. over pixels surrounding the point locations is performed. Irradiance values obtained after spatial averaging have 3-h temporal resolution, which are to be interpolated to GHI values with 1-h temporal resolution.

Hereafter for point forecasts, two different techniques are used for temporal interpolation. In technique1, direct linear interpolation technique is employed to reduce temporal resolution of 3-hourly NWP model GHI values to 1 h, while in technique2, interpolation was performed on 3-hourly CSI (Kt*) values (using the Bird clear sky model) to obtain hourly Kt* values [14]. Then, hourly Kt* values are multiplied by hourly clear sky model values to obtain hourly NWP mainly adopted to account for the typical is mainly adopted to account for the typical diurnal pattern of irradiance. For multiple-site analysis, similar procedures are repeated with averaged forecasted NWP model GHI values from all point locations under study. The forecasted GHI series, obtained by employing these two techniques, are validated against acquired GHI datasets from the SRRA stations in Rajasthan [75, 122].

A comparison of the two techniques was performed by obtaining different measures of accuracy [12]. As can be seen in Table 3, a reduction in rRMSE for technique2 is observed, which signifies an improvement of forecast accuracy for the second approach. But the forecasts from technique2 could be further improved by using the MOS approach that removes systematic bias errors which depend on the time of the day and on sky conditions, etc. So, in technique3, the intention was to model the bias as a function of hourly clear sky index (Kt*) and the cosine of solar zenith, using the measurement data obtained from SRRA stations for the first 15 days of each month in 2014 as training set and validate the model using test set which comprises of measurement datasets for remaining days. Curve fitting parameters are evaluated and a good correspondence between the modelled and measured bias values is observed. Then, the improved forecast is obtained by subtracting the modelled bias values from the forecasts obtained using technique2 [81, 84].

Figure 10 shows the forecasted GHI series obtained from technique3 with the actual GHI values measured at Aburoad for a single-site analysis, and Fig. 11 compares the forecasted GHI series with averaged GHI values for a multiple-site analysis. After calculating error measures for technique3 (Table 3), it is observed that the forecasted GHI series evaluated using technique3 has a higher accuracy compared to the previous two techniques. Again, MOS updated forecasts with multiple-site analysis has a lower uncertainty compared to MOS updated forecasts with single site/local area analysis. For single-site analysis, the forecast accuracy is very good during clear sky conditions. However, the level of uncertainty increased during cloudy conditions. For MOS updated GHI with wide area analysis, the uncertainty reduced drastically because of smoothing due to spatial averaging reducing the deviation between forecasted and mean of measured GHI values (Table 3).

From Fig. 12, it can be observed that the least rRMSE value of around 9% is achieved for forecasted GHI series from technique3 with wide area analysis. The least percentage rRMSE signifies highly accurate forecasted GHI series, and this result is in agreement with the accuracy values observed in [90]. Figure 13 depicts relative MBE values obtained for different day-ahead forecasting techniques adopted. Very small rMBE values are obtained for MOS techniques with bias removal (Table 4).