Chief Data Officer / China
Decisions that cost revenue or capacity to get right
- Electricity supplied
- 9.4 TWh
- 2025; 1,860 commercial and industrial customers
- Public charging connectors
- 11,200
- Across four cities in China
- Data office colleagues
- 58
- Analytics, engineering, governance and personal information
Decisions & consequences
01Reading our largest customers' own meters before buying a price forecast
In 2023, our day-ahead load declarations for the 60 largest customers, who used 41 per cent of our volume, rested on two weak inputs. Grid metering data reached us a day or more after the trading day, and production plans arrived by email when customers remembered them. Settling the resulting differences at real-time prices had cost RMB38 million in 2022.
Our head of trading, Haoming Jie, wanted an external weather and price-forecasting subscription at RMB3.4 million a year. He believed price error drove most of that cost. The finance director questioned placing equipment at customer sites when contracts renewed annually. I proposed metering gateways reading each customer's own sub-meters every fifteen minutes, at RMB5.8 million to install and RMB1.2 million a year to run.
A back-test of 2022 attributed about two-thirds of the deviation cost to load error at large sites, and that persuaded the general manager. We postponed the subscription for a year to fund the gateways. Of the 60 customers, 44 accepted written data-use terms: forecasting, settlement checks and their own energy reports only, with deletion ninety days after a contract ends. Sixteen declined. Two customers with gateways moved to another retailer in 2024, stranding RMB260,000 of equipment.
At the 44 sites, weighted day-ahead load error fell from 8.7 per cent to 4.6 per cent in 2024. Portfolio deviation cost fell to RMB27 million, but spot prices were also calmer that year; our own analysis credits the gateways with about RMB6 million of the reduction, and I have not claimed more. Gateway customers now receive a monthly note showing how their data changed the forecast, and we took the price subscription in 2025 after testing it against our results.
I would change two things. I would write the data terms into the retail contract at renewal instead of collecting separate letters. I would also have trialled the price service in parallel, rather than presenting the board with a choice between the two.
Selling station aggregates instead of drivers' charging histories
In early 2024 a vehicle manufacturer offered RMB3.6 million a year for session-level charging records: time, station, energy delivered and the vehicle identifier some of our chargers receive during connection. It wanted to study battery behaviour and to contact owners about its services. At that point we still kept every identifiable session indefinitely, about 260 million records, because nobody had decided otherwise.
Our commercial director, Peiran Shao, argued that removing phone numbers would make the records anonymous and that the revenue would fund 300 new connectors. I disagreed. A vehicle identifier with times and places still describes where a person goes, and Chinese personal-information law treats whereabouts as sensitive. Providing those records to another company would need each driver's separate consent and an impact assessment, and few drivers would have expected their charging history to leave our app.
The general manager accepted my recommendation after legal colleagues reviewed the analysis. We declined the session-level sale and offered station-hour aggregates of occupancy and energy, suppressing any cell with fewer than ten sessions. The manufacturer paid RMB1.1 million for the first year. We gave up RMB2.5 million of annual revenue, and Peiran's connector plan for that year was reduced to 180.
The offer also exposed our own retention. We set identifiable session records to twenty-four months, enough for finance and complaint handling, deleted plate numbers after thirty days and aggregated everything older. Identifiable sessions held fell from 260 million to 31 million by mid-2025. An external review of the app's personal-information practices that year raised two minor points about consent wording, which we corrected.
What I would do differently: the retention clean-up should have come before any buyer arrived, not because of one. I would also have proposed an opt-in research panel from the start. We piloted one in 2025, and 8,400 drivers joined on clear terms, giving the manufacturer better evidence than the records it first asked to buy.
Locking out 49 connectors and limiting 183 before the Spring Festival peak
In January 2025, three weeks before the Spring Festival travel period, my governance team linked charger telemetry to the maintenance contractor's work orders for the first time. Alarms and work orders had been recorded separately since 2021. The link found 49 connectors with repeated insulation-monitoring alarms that technicians had closed as resets with no fault found, and 183 with repeated over-temperature current reductions. Many were at our busiest sites, including two intercity service areas.
The safety manager, Xiaodan Rui, wanted all 232 locked out until inspected. The operations director argued that most alarms were nuisance faults and that removing capacity before the holiday would cause queues and complaints. The data could not settle the disagreement. It showed patterns rather than confirmed defects, and for 31 of the connectors the telemetry was too incomplete to judge either way.
We made a joint recommendation and the general manager decided. The 49 insulation-alarm connectors were locked out immediately. The 183 over-temperature connectors stayed in service with current capped at a level set by the engineering team, and were inspected in order of alarm frequency. The 31 with incomplete telemetry went to the front of that list. Xiaodan disagreed with keeping capped connectors live and recorded her view in the decision paper. We lost about RMB1.9 million of charging revenue over the holiday weeks, and queues at the service areas drew complaints.
Inspection found damaged cable insulation or water ingress at 17 of the 49 locked-out connectors and worn contacts at 22 of the 183. No thermal incident occurred. The contractor's agreement now requires a cause code before a work order can close, and the link between alarms and work orders runs every night.
The data had existed for four years. I had given trading analytics priority and left these safety records unread, and that delay was my decision. I now review each quarter which safety-relevant datasets nobody in my office has examined, and I report that list to the risk committee.