Learnings

Outcomes

The outcomes of the project are as follows;

• We have a better understanding of smart meter data availability and the processes required to gather data.
• We have a better understanding of the response levels from smart meters to being reconfigured to and from 1-minute voltage data measurement.
• We have gained an understanding of the general level of data inaccuracy in the NGED asset register, CROWN, for feeder and substation association and the phase data within the NGED Geographical Information System, Electric Office (EO).
• We have improved our ability to gather more targeted aggregated data as a result of revising the data privacy agreement.We have established the actual phase connectivity for 8,790 customers linked to the 46 distribution substations within the test network in the Milton Keynes area. 
• We have established the actual feeder connectivity for 125 customers from the feeder survey.
• We have captured data about LCTs that were visible from the roadside which can be used to assist with LCT validation.
• We have captured photographs of the inside of customer meter boxes which can be used in future projects e.g., AI image processing.
• We have developed and validated methods for phase and feeder identification with indicative accuracy metrics.  The way in which these can be set up for BAU is currently under discussion.
• We have established the best of four proposed methods for estimating load profiles for LV feeders and for distribution substations.
• We have identified potential methods for determining the locations of LCTs. The locations of PV installations can be determined with high confidence using export demand data.
• Visual validation for domestic PV has been attempted using free to access on-line resources.
• The learning from the project has been shared with stakeholders and published on the NGED website.

Lessons Learnt

Data Quality and Availability

As SMITN was one of the first projects making substantive use of smart meter data, some limitations were discovered around the retrieving the data. The project also uncovered some previously unknown issues with NGED’s legacy data. 

• Smart meter MPAN coverage - Around 47% of the 8,809 MPANs in the trial area had smart meters installed
• Half-hourly voltage data availability - Around 2,200 devices gave a response out of a population of 4,140, which gives a success rate of 57%.
• 1 minute voltage data availability - When the 4,140 smart meters were sent a message to configure them to record 1-minute voltage data, only 1,800 gave a positive response, i.e. under 50%
• Clock synchronisation - The timestamps in the smart meter voltage data were generally well-aligned to the substation monitoring data with any offsets being mostly within a few minutes. These time offsets could be corrected by correlation method relative to the substation monitoring data.
• Voltage measurement interval synchronisation - Separately from the issue of clock synchronisation, many of the smart meters used randomised starting times for the voltage measurement intervals. Half-hourly voltage data could be recorded starting at any time within a clock half-hour, rather than at 00 and 30 minutes past the hour.
• Data provision timescales - the complexity of data extraction from smart meters had been underestimated but now that the required reports are created and approved that should not be an ongoing issue.
• Smart meter uptake vs LCT uptake - Customers with LCTs are more likely to have a smart meter than customers who do not.
• Visibility of tap change impacts - The correlation between busbar voltage and tap changer position is not as clear as expected which meant that tap change operations could be used as a way to filter out related voltage changes on smart meters.
• Monthly consumption data outliers - The monthly consumption data being provided by the smart meters for individual customers showed a number of customers having unusual values. Therefore, some data cleansing and sense checking was required for this data item.

Phase Identification 

• Technique Suitability – both clustering and correlation techniques provided good results for phase identification with better results being seen where monitoring data was available for the substation. Errors in CROWN data resulted in data from different feeders being compared together which reduced overall accuracy.
• Voltage step change vs absolute voltage – creating a data series of voltage step changes provided better results than using the absolute voltage data
• Clustering method - Hierarchical clustering worked better than K-means.
• Looped Services – this provided an additional way to validate results as they should be the same phase. 
• Time resolution – 1 minute voltage data gave better results than 30 minute voltage data

Feeder validation

• Iterative approach – Phase identification and substation profile creation can highlight customers on the wrong feeder so an iterative approach between different methods may be useful.
• methods – as for phase identification, using voltage clustering and correlation can be used to determine a customer feeder.
• Voltage step change vs absolute voltage – creating a data series of voltage step changes provided better results than using the absolute voltage data
• Time resolution – 1 minute voltage data gave better results than 30 minute voltage data

LCT identification

• Smart Meter uptake - Customers with LCT are more likely to have a smart meter than customers who do not.
• Voltage alarms – These were assessed for their potential to indicate unregistered LCT. It was found that while customers with LCT were more likely to have voltage alarms the data was not able to provide conclusive relationships.
• PV detection - Half-hourly active export readings are a good indicator of PV installations. In the test network area 94% of known PV installations can be identified by having non-zero export data. As expected, PV Export peaks around midday during the summer months.
• PV validation – Two thirds of the premises identified as likely to have unregistered PV were able to confirm a PV installation using free-to-use satellite and street view imagery.
• EV detection – consumption data -  Many of the energy consumption features for customers with and without EV chargers are similar with large areas of overlap. However, machine learning algorithms for binary classification trained using multiple features of the 12 months of demand data, were able to correctly identify customers with EV chargers. A Weighted Support Vector Machine method was found to give the best detection rate while maintaining a false positive rate below 10%. The summer consumption and the maximum consumption features have higher impact on the models.
• EV detection from voltage data -  An algorithm searched for significant voltage drops with high duration. This had high levels of false positives as other large loads with long duration could cause the same voltage impacts as EVs. However, we could use this approach to validate and reduce the number of the unlabelled properties that have been predicted as having an EV from the “EV detection using monthly demand data” approach. A Machine Learning model could identify which properties have EVs based on frequency of drops, average drop lengths, and percentage of other Smart Meters on the same feeder that the algorithm detects drops relative to.

Substation profiles            

• Network model accuracy - Problems with customers being incorrectly associated with the wrong LV feeder or substation can introduce inaccuracy into the estimates for LV or substation profiles. In one example a half-hourly metered customer was attributed to the wrong LV feeder which resulted in inaccurately high and low estimates on the real and assumed LV feeder.
• Profile methods - Given that the aggregated demand for smart meters and for traditionally half-hourly meters MPANs is known, the accuracy of the estimated load profiles depends on the process used to model conventional non-half-hourly meter demands.
• Full vs abridged Elexon method - Load profiles for Non Half-Hourly (NHH) meters calculated considering daily temperature compensation and Time Pattern Regimes (TPR) have a higher accuracy than those using the set of 15 generic Elexon profiles.
• Scaling up - An alternative approach to estimating the demand of NHH meters is to scale the demand from smart meters located on the same feeder, treating the domestic smart meters as a proxy for the demand of the neighbouring domestic NHH meters.  This was found to be less accurate than using the Elexon data.
• Daily peak estimation - The profile estimation methods for LV feeders calculated daily peaks within 5% to 20% of measured values. This range indicates the Root Mean Square (RMS) over the test duration (approx. 9 months) of the daily errors.