1/14
Machine learning-based energy management for a microgrid with renewables and hybrid storage
Title Card
Welcome to my presentation on the design and analysis of a machine learning-based energy management system for a microgrid with renewables and hybrid energy storage
Outline
I will be covering the aims and literature, the methodology, results, findings and conclusion, with heavy focus on methodology and results.
Problems and aims
A microgrid system with PV, a hybrid storage system and a backup diesel generator requires an energy management system with the task of maintaining low cost when grid connected and providing reliable supply in stand-alone mode.
Rule-based EMSs work but require manual tuning.
This project tests whether machine learning can outperform these models.
Literature review
Three findings from my research shaped my design.
Cycle charging beats load following, thus charging the battery from the generator cuts diesel hours and fuel use because the generator avoids running at light loads.
Rule-based EMSs cannot adapt. Threshold rules are precise and auditable, while reinforcement learning can learn dispatch decisions directly from training.
Additionally, safely layers are required as learned actions must stay within physical limitations.
Methodology 1: the microgrid
The PV array: a surrogate model with P&O maximum power point tracking.
Hybrid storage: a 2.75kWh battery and a 3.8kW supercapacitor
A backup diesel generator at 3kW, started and stopped by battery state-of-charge thresholds
The surplus handled by a 4kW geyser and 6kW dump load, in addition to the grid connection.
Each component was modelled in Simulink and assembled into a power-balance hub and a faster reduced-order model used for training.
Methodology 2: the step and the agent (Slide 6.1)
The same calculation every simulated 60 seconds, five stages:
PV and load: irradiance and temperature give PV power and a load with one fixed daily shape.
Storage split: power-frequency decoupling, with the supercapacitor taking the high-frequency component and the battery the low-frequency component.
EMS: generator ON and OFF thresholds on battery state of charge.
Balance and cost: power balance, fuel, starts, grid and unserved energy.
Reward: cost saved against the baseline on the same day.
Slide 6.2
The agent is a Soft Actor-Critic: a small neural network that sees 17 measurements of the system and, every 5 minutes, decides how far to shift the generator鈥檚 ON and OFF thresholds.
Training: 1000 synthetic days, each randomly sunny or cloudy with a random starting battery level. It is the tested on 60 unseen days, 30 of each.
Methodology 3 of 3: slide 7
The baseline is a cycle-charging rule based on Ameen et al. for cycle charging, with threshold rules in the style of Huo et al.: the generator starts at full power when the battery falls below a minimum threshold, recharges it while serving the load, and stops at a maximum threshold.
Slide 7.2
The agent is compared with three rule-based controllers.
B0 is the untuned rule: the generator starts when the battery falls to 25% and stops at 60%.
B1 is the baseline. Its two thresholds were tuned by trying every combination and keeping the cheapest, giving 30% and 80%.
B1-star repeats that search for each battery size.
The agent is B1 with learned shifts to those thresholds.
Slide 7.3
The main test is 60 unseen days, each run with both the agent and B1, alongside four reference scenarios and 20 days in the full Simulink model. Supporting studies cover prices, battery size, the supercapacitor, the dump load and a network load flow.
Results 1: Validation(Slide 8)
PV surrogate error 0.32%, P&O tracking efficiency 0.9987, battery delivers 0.999 and 1.001 of commanded power, reduced model within 2% on grid cost and 5% on fuel
Network: PowerFactory load flow 0.999 to 1.002 p.u., losses below 0.05 kW, maximum loading 80.1%
these checks are why you can trust what follows
the reduced model understates unserved energy for B0, and I did not isolate the cause
Results 2: slide 9
60 unseen stand-alone days, cost per day against B1
Sunny days: R5.43 cheaper, better on 27 of 30
Cloudy days: R8.97 more expensive, worse on all 30
Overall: R1.77 more expensive per day (standard error R0.99), so no improvement
Why: on cloudy days the agent lowers the OFF threshold (about 0.72 against 0.77 to 0.80), so generator runs are shorter
The battery drains again quickly, so starts rise from 6.0 to 7.9 per day
Land: the extra starts and fuel cost more than the agent saves
Results 3: tuning and battery size(Slide 10)
Battery sweep, thresholds re-tuned for each size: 2, 4 and 8 times the 2.75 kWh battery lowers cost by R13.6, R25.7 and R43.1 per day
Tuning B0 into B1 is worth R11.51 per day, about twice the agent's best gain
Supercapacitor: ramp rate down 98%, peak battery current down 7 to 17%
Geyser absorbs 58% of surplus PV on the sunny stand-alone day and 99.9% on cloudy stand-alone.
Say: the agent was trained at 2.75 kWh only, so the sweep used the rule controllers
Overall findings(Slide 11)
The models can be trusted: PV surrogate within 0.32%, reduced model about 2% from the hub on grid cost
The agent does not beat the baseline: R1.77 more expensive per day, R5.43 cheaper on sunny days, R8.97 more expensive on cloudy days
The sign depends on two prices: the overall result flips only if a start costs under R0.26 (base R2) or stored energy is valued under R3.70 per kWh (base R8.25)
Tuning and battery size matter most: R11.51 per day from tuning, R43.1 per day from a 22kWh battery.
Add on point 3: the cloudy-day loss stays even with no credit at all
Conclusion(Slide 12)
Deliverables one to five were met: the literature reviews are complete, and the microgrid and utility grid models were built, verified and combined into a working simulation platform.
Deliverable VI, partly met: an ML EMS was built and validated with cost as its objective. Saves R5.43 per day on sunny days, R1.77 per day more expensive overall. In grid-connected mode it matches the baseline, with no extra grid-cost saving
Deliverable VII, partly met: resources varied (sunny and cloudy days, four scenarios, battery 2.75 to 22 kWh, prices), load held at one fixed shape, agent under varying load untested
Next steps: a hybrid gate (agent on sunny days, baseline on cloudy days), retrain at larger batteries, test low, medium and high load and a grid-connected agent with battery power action, to cut grid cost