Context
Whether operating the network day-to-day or planning its evolution, Distribution Network Operators (DNOs) need to accurately estimate the power flowing through their MV/LV transformers. These estimations help identify potentially overloaded transformers, allowing DNOs to consider replacement or relieve them through local network restructuring. Conversely, a transformer identified as consistently under-loaded can be reallocated to another public distribution substation, increasing its utilization rate and improving investment efficiency. Knowing the transformer’s load factor also helps verify that connecting a new user, or increasing an existing customer’s subscribed power, won’t lead to an overload situation.
In this context, the deployment of smart meters is a pivotal event. The new data produced by these meters offers an ideal raw material for improving load estimation methods within the distribution network, particularly for public distribution transformers. However, metering data is sensitive, and its use is regulated, which constrains how DNOs can utilize this data.
Problem Statement
Various solutions exist for directly measuring the maximum power flowing through MV/LV transformers, ranging from manual maximeter readings and resets to the installation of a Remote Terminal Unit (RTU) integrated into SCADA for continuous remote monitoring of the transformer’s status. However, these solutions incur costs, and are therefore not systematically adopted by DNOs. Furthermore, direct measurement only provides an estimate of past maximum load, whereas planners are generally more interested in the maximum load likely to occur in the future with a certain level of risk. Thus, it’s relevant to explore statistical models to estimate MV/LV transformer load, either as a complement to or in place of direct measurement.
A common statistical approach to estimate the peak power of an aggregation of distribution network users is to sum their subscribed powers and then multiply by a “diversity factor.” This factor’s value decreases as the number of users increases. NF C14-100 standard, among others, provides the following coefficient table:
| Number of users downstream of the considered section | Coefficient |
|---|---|
| 2 to 4 | 1 |
| 5 to 9 | 0.78 |
| 10 to 14 | 0.63 |
| 15 to 19 | 0.53 |
| 20 to 24 | 0.49 |
| 25 to 29 | 0.46 |
| 30 to 34 | 0.44 |
| 35 to 39 | 0.42 |
| 40 to 49 | 0.41 |
| 50 and above | 0.38 |
This was the method GreenAlp, the Distribution Network Operator for the city of Grenoble and 23 other municipalities in Isère (France), had used until now. However, the limitations of this method led GreenAlp to approach Roseau Technologies to develop a more precise approach.
Achievements
Two supervised learning prediction models were proposed to GreenAlp. The first method involved retaining the current methodology and only updating the numerical values of its parameters, meaning adjusting the diversity factor table to better reflect the behavior of GreenAlp’s customers. The second solution involved using a new, more precise prediction formula, based on an additional descriptor: the total energy consumed by a customer over the 30 days of their maximum consumption. This indicator can be calculated in practice using daily consumption indices, which DNOs are authorized to collect freely.
A set of load curves at 10-minute intervals (for customers >36 kVA) or 30-minute intervals (for customers <=36 kVA), recorded using the smart metering infrastructure, were available for the study. For customers <=36 kVA, these load curves can only be legally collected by the Network Operator under specific exemptions, and were therefore only available for approximately 10% of customers and for a duration of about one year. This constraint mandated using this data only for the training phase, excluding the prediction phase.
Roseau Technologies carried out the following work:
- Identify, collect, and consolidate the various available input data (load curves, but also mapping data, SCADA data, etc.).
- Define and evaluate different prediction methods, and propose to GreenAlp a selection of two methods chosen for their simplicity and prediction performance.
- Calculate the new values of the diversity factors, adjusted based on the extract of load curves, replacing the standard coefficients previously used by GreenAlp.
- Similarly, calculate the numerical parameters of the improved prediction method through learning.
- Evaluate the prediction performance of each of the two methods.
- Finally, provide an improved estimation of the load factor for all public distribution transformers in the city of Grenoble.
Impact
The project immediately enabled GreenAlp to detect several potentially overloaded MV/LV transformers. These transformers were then equipped with mobile measurement devices to confirm the diagnosis and, if necessary, consider a corrective action such as transformer reassignment. The project also helped identify some lightly loaded transformers and reallocate them to high-load areas. Finally, it led GreenAlp to update the calculation method used in its network sizing software.
Publications
This work was the subject of a scientific publication, presented at the CIRED 2025 conference.
