Source-linked AI summary
Energy management of small-scale PV-battery systems: A systematic review considering practical implementation, computational requirements, quality of input data and battery degradation
Donald Azuatalam, Kaveh Paridari, Yiju Ma, Markus Förstl, Archie C. Chapman, Gregor Verbič
TL;DR
Residential HEMS strategy comparisons have not holistically assessed practical implementation alongside forecast uncertainty and modelling assumptions. This paper compares seven strategies and analyses their techno-economic performance under differing forecasts, computational demands, and battery degradation. It finds that greater strategy sophistication does not necessarily improve PV-battery performance or economic viability, and cautions that results may not hold in practice under the modelling assumptions used.
Problem
Prior assessments did not holistically compare HEMS strategies while accounting for forecast errors, modelling assumptions, and practical implementation.
Source: §2.2 · p. 4, §8 · p. 16Method
The study compares seven HEMS strategies and evaluates their techno-economic performance under forecast uncertainty, computational limitations, battery degradation, and modelling assumptions.
Source: §8 · p. 16Results
More sophisticated strategies do not necessarily improve PV-battery performance or economic viability, while outcomes depend on input-data quality and battery degradation.
Source: §8 · p. 16Takeaways & Limitations
When accurate PV and demand forecasts are unavailable and no additional battery controllers compensate for errors, the authors suggest simple heuristic or machine-learning strategies over optimisation-based methods.
Source: §8 · p. 16Simulation findings may not hold in practical implementation because of strategy-specific modelling assumptions, and the authors plan hardware experiments to confirm them.
Source: §8 · p. 16
Abstract
from arXiv · showhide
The home energy management problem has many different facets, including economic viability, data uncertainty and quality of strategy employed. The existing literature in this area focuses on individual aspects of this problem without a detailed, holistic analysis of the results with regards to practicality in implementation. In this paper, we fill this gap by performing a comprehensive comparison of seven different energy management strategies, each with different levels of practicality, sophistication and computational requirements. We analyse the results in the context of these three characteristics, and also critique the modelling assumptions made by each strategy. Our analysis finds that using a more sophisticated energy management strategy may not necessarily improve the performance and economic viability of the PV-battery system due to the effects of modeling assumptions, such as the treatment of uncertainties in the input data and battery degradation effects.
1. Introduction
Residential HEMS coordinate distributed energy resources to reduce electricity costs and support returns on substantial PV-battery investments, but performance depends on tariffs, forecasts, and strategy practicality. This paper compares seven strategies and assesses their techno-economic performance against forecast uncertainty, computational limits, battery degradation, and modelling assumptions.
1.1. Background and motivation
- HEMS schedule distributed energy resources and household appliances to reduce electricity costs while maintaining acceptable thermal comfort.Owners also seek timely recovery of PV-battery investment, with financial benefits depending on strategy performance and the retail tariff.Source: §1 · p. 1, §1.1 · p. 2
- Self-consumption maximisation stores excess PV generation for later use, while time-of-use arbitrage pre-charges batteries from the grid during cheap off-peak periods.The two rule-based heuristics shift energy use but do not explicitly minimise electricity cost.Source: §1.1 · pp. 2–3
- The comparison includes optimisation-based MILP and dynamic programming, alongside neural-network policy-function approximations trained on optimisation outputs and rule-based heuristics.The strategies are assessed in light of their assumptions, available information, and practicality.Source: §1.1 · p. 3
1.2. Contributions
- Earlier work did not provide a detailed holistic comparison of strategy performance under forecast errors and modelling assumptions, including practical implementation of advanced optimisation approaches.This study extends prior review and preliminary comparison work to examine these facets.Source: §1.2 · p. 3
- The study models and compares seven HEMS strategies using real customer load traces and Sydney-region tariffs, considering forecast uncertainty, computational limits, and battery degradation.Its analysis also examines how modelling assumptions and practical implementation affect the results; the best strategy often depends on input-data quality and forecast accuracy.Source: §1.2 · p. 3
2. Literature review
The literature reviews HEMS performance across cost savings, forecasts, data resolution, computational feasibility, tariffs, and battery degradation. It highlights trade-offs among accuracy, uncertainty, model complexity, speed, and practical assumptions.
2. Literature review
- The review spans six factors affecting HEMS performance: cost savings, load and PV forecasting, temporal resolution, computational feasibility, retail tariffs, and battery degradation.The paper compares PV-battery systems using optimisation-based strategies with simple heuristics under the same inputs and assumptions.Source: §2 · p. 3, §2.1 · p. 4
2.2. Load and PV generation forecasting
- Perfect foresight is common in prior optimisation studies despite being unrealistic, and this study uses persistence demand forecasts to analyse forecast-error effects on optimisation methods and economic indicators.Day-ahead household demand is harder to predict than PV output; the cited persistence-forecast study reported a 4.4% SCR reduction versus perfect forecasts when using a PI controller.Source: §2.2 · p. 4
2.3. Temporal resolution
- Half-hourly PV and demand data are used because storage mitigates the effect of lower temporal resolution on self-consumption errors in PV-battery systems.Prior findings cited in the review indicate that PV-only systems need 15-minute resolution for reasonably accurate results.Source: §2.3 · p. 4
2.4. Modelling complexity and computational feasibility
- Accounting for uncertainty and nonlinear battery and inverter operation can improve model realism but increases modelling complexity and computational burden.DP incorporates uncertainty and nonlinearities for high-quality solutions, whereas MILP uses linearised battery operation and deterministic forecasts for faster, lower-quality solutions; PFA shifts computation offline for faster online execution.Source: §2.4 · pp. 4–5
2.5. Retail tariff
- Reported tariff profitability varies with tariff figures, peak periods, and customer response; this study considers ToU tariffs because the authors state that they maximise PV-battery cost savings.Prior studies cited in the review report differing outcomes across flat, ToU, wholesale, and critical-peak tariffs.Source: §2.5 · p. 5
2.6. Battery degradation
- Battery aging is often omitted from operational studies, and the reviewed degradation models include assumptions that charging and discharging age batteries equally.The study uses SimSES for its flexibility and accuracy in techno-economic analysis of lithium-ion batteries.Source: §2.6 · p. 5
3. Methodology
The methodology evaluates residential PV-battery strategies using a common techno-economic framework and specified system, tariff, customer-data, and battery assumptions. Scheduling outputs feed financial and degradation analyses.
3.1. Framework and system topology
- Strategy outputs—grid power, battery charge/discharge power, and battery state of charge—feed financial and degradation analyses using savings, IRR, SOH, full equivalent cycles, and expected lifetime.Demand, PV output, DER, and tariff data are inputs to the scheduling algorithm, evaluated under perfect and persistence forecasts.Source: §3.1 · p. 5
- The model assumes separate PV and battery inverters, while acknowledging a common-inverter topology as another configuration used in practice.PV can serve load directly or be exported; battery and grid power can flow in both directions.Source: §3.1 · p. 5, Figure 2 · p. 6
3.2. Electricity retail tariffs
- The study uses residential retail charges from Origin Energy in New South Wales and models ToU pricing periods alongside feed-in tariffs.The tariff prices include electricity costs, retailer risk-management and service fees, and network charges.Source: §3.2 · pp. 5–6, Figure 3 · p. 6
3.3. Load and DER data
- Demand and PV profiles come from 52 clean Ausgrid customer records at half-hourly resolution, covering three years and PV systems ranging from 3 to 10 kWp.The dataset's PV inverter efficiency is already accounted for, so simulations set it to 1; battery sizes depend on PV size.Source: §3.3 · p. 6
3.4. Battery operating model
- Battery efficiency is modelled as a constant 91% charge/discharge efficiency, equivalent to 84% round-trip efficiency, although the authors note this can yield infeasible results for linear-battery strategies.The chosen value reflects cited estimates of typical lithium-ion battery efficiency including losses.Source: §3.4 · pp. 6–7
4. Energy management models and simulation assumptions
The paper describes seven energy-management strategies and the simulation assumptions used to assess them, with strategy details summarized for a self-contained account.
4. Energy management models and simulation assumptions
- The section summarizes the seven energy-management strategies and their simulation assumptions, while directing readers to the cited references for fuller details.The authors state that space constraints limit the strategy descriptions to summaries.Source: §4 · p. 7
4.1. Home energy management strategies
The paper compares rule-based, optimisation, and policy-function strategies for residential PV-battery management. Their trade-offs reflect modelling choices, including uncertainty treatment and battery operating constraints.
4.1. Home energy management strategies
- The seven strategies span self-consumption and tariff heuristics, MILP, DP, and customer-specific or generic-day-type policy function approximations.The paper lists SCM, ToUA, SCM+ToUA, MILP, DP, PFAS, and PFAG.Source: §4.1 · p. 7
4.1.1. Self-consumption maximisation (SCM)
- SCM prioritises serving demand with PV, then directs surplus to battery charging or grid export.The paper identifies SCM as a baseline strategy used by retailers and battery suppliers.Source: §4.1.1 · p. 7
4.1.2. Time-of-use arbitrage (ToUA)
- ToUA pre-charges the battery with cheap grid power for higher-price periods, and is beneficial only for some customers under time-of-use tariffs.Its suitability depends on the customer’s load profile.Source: §4.1.2 · p. 7
4.1.3. Self-consumption maximisation with time-of-use arbitrage (SCM+ToUA)
- SCM+ToUA uses SCM as its baseline and applies tariff arbitrage only when a perfect forecast indicates low PV generation.Source: §4.1.3 · p. 7
4.1.4. Mixed integer linear programming (MILP)
- MILP explicitly minimises electricity cost over a decision horizon with fixed tariff prices, but neglects stochasticity and linearises the battery SOC transition.The authors say neglecting stochasticity results in a lower-quality solution.Source: §4.1.4 · pp. 7–8
4.1.5. Deterministic dynamic programming (DP)
- Nonlinear battery and inverter operation and maximum charging rates give DP higher-quality solutions than the other strategies.The DP model is formulated as a Markov decision process; this paper uses its deterministic version.Source: §4.1.5 · p. 8
- DP formulates HEMS as a Markov decision process and uses value iteration to evaluate expected future discounted cost for a policy.In this study’s deterministic version, the effects of the exogenous random variable are neglected.Source: §4.1.5 · pp. 8–9
4.1.6. Policy function approximation using customer specific day types (PFAS)
- PFAS trains a customer-specific neural network offline on historic PV and demand data using MILP’s optimal policy, then predicts grid power online.A control filter is applied because the PFA formulation does not itself consider battery-operation constraints.Source: §4.1.6 · p. 9
4.2. Demand and PV persistence forecasting
The study tests persistence forecasts to assess energy-management performance with imperfect PV and demand predictions. Demand uses the same weekday profile from the prior week, while PV prediction includes a bounded random error.
4.2. Demand and PV persistence forecasting
- The persistence forecasting setup evaluates energy-management strategies with imperfect demand and PV-generation forecasts.Source: §4.2 · p. 9
4.2.1. Demand ˜Pd prediction
- Demand for each forecast time step is set to the demand at the same time one week earlier.The prior-week weekday profile is used to reflect seasonality in consumption.Source: §4.2.1 · pp. 9–10
4.2.2. PV ˜Ppv prediction
- PV generation is forecast using the previous day’s output with a uniform random error that limits deviation from actual output to 10%.Source: §4.2.2 · p. 10
5. Economics
The economic assessment specifies system costs and electricity-price assumptions, then uses annual cost savings and internal rate of return to assess PV-battery viability. Inflation is included in the return calculation.
5.1. Cost parameters
- The analysis assumes electricity prices rise by 3% annually for 20 years, noting that omitting inflation would produce relatively lower IRR values.Initial investment is defined as the total PV and battery system cost.Source: §5.1 · p. 10
5. Economics
- The financial assessment uses annual cost savings and internal rate of return (IRR) as its two economic viability indicators.Source: §5.2 · p. 10
5.2. Financial indicators
- IRR is the discount rate at which NPV is zero; the analysis levelizes annual savings and adjusts the return calculation for electricity-price inflation.Source: §5.2 · p. 10
6. Battery degradation study
The degradation study uses SimSES to estimate battery aging, capacity degradation, state of health, and expected lifetime from operating and aging inputs.
6. Battery degradation study
- SimSES uses yearly battery charge/discharge power, SOC, and aging parameters to estimate SOH after 20 years, annual equivalent full cycles, and lifetime at 80% SOH.Inputs include depth of cycling, C-rate, Ah-count, SOC limit, and temperature; the model combines calendar and cyclic aging.Source: §6 · p. 10
7. Results and analysis
The results compare scheduling, economic performance, computation time, and battery aging across energy-management strategies. They show that modeling assumptions and forecast quality can materially affect simulated feasibility and practical performance.
7.1. Scheduling results
- DP schedules cost-based charging and discharging with a nonlinear battery model; ToUA reserves 30% of maximum SOC, while PFAS and PFAG produce different MILP-trained schedules.ToUA’s reserve is intended to provide energy security when day-ahead forecasts indicate low PV generation; PFAS is customer-specific, whereas PFAG uses a cluster representative.Source: §7.1 · pp. 11–12, Figure 10 · p. 13
- Linear battery models can make MILP and SCM schedules infeasible under nonlinear battery behavior, potentially overestimating HEMS economic performance.The SOC comparison illustrates this constraint violation.Source: §7.1 · p. 11, Figure 8 · p. 12
7.2. Economic viability
- Under imperfect forecasts, SCM and SCM+ToUA save more than MILP and ToUA, while PFAS and PFAG are more robust than optimization approaches.With perfect forecasts, MILP has the highest median annual cost savings among the compared heuristic strategies; simulations favor MILP over DP, but the authors expect DP to outperform in practice because it models battery and inverter nonlinearities.Source: §7.2 · p. 12, Figure 11 · p. 14
7.3. Computation time
- SCM solves a customer’s daily problem in under 0.006 seconds, compared with 0.084 seconds for MILP and about 30 seconds for DP.PFAS takes 0.0118 seconds to execute, while PFAG takes nearly 0.6 seconds; reported computational-speed comparisons assume linear battery models.Source: §7.3 · p. 14, Table 6 · p. 14
7.4. Battery degradation results
- DP has the highest expected median battery lifetime at 20 years, versus 11–13 years for other strategies, and a median SOH of 80%; ToUA performs best in the SOH category.Source: §7.4 · p. 15
- Optimization strategies produce more full equivalent cycles, while DP has the least average cycle depth and the strongest reported state-of-health outcome.PFAS shows slightly better aging results than MILP, and its smoother schedule is associated with fewer full equivalent cycles; the authors advise considering degradation when choosing a strategy.Source: §7.4 · pp. 14–15, Figure 12 · p. 15
8. Discussion and conclusions
The paper compares seven energy-management strategies and finds that greater sophistication does not necessarily improve PV-battery performance or economic viability; practical results depend on input-data quality, battery degradation, and modeling assumptions.
8. Discussion and conclusions
- Greater strategy sophistication does not necessarily improve PV-battery performance or economic viability, as input-data quality, battery degradation, and modeling assumptions affect outcomes.The authors caution that simulated results may not hold in practical implementation.Source: §8 · p. 16
- With a linear battery model, well-tuned heuristics can approach MILP cost savings; DP is closer to practice with a realistic battery model and outperforms other strategies on aging.SCM+ToUA can slightly improve on SCM, the baseline heuristic.Source: §8 · p. 16
- Without accurate PV-demand forecasts or compensating battery controllers, simple heuristic or machine-learning strategies may be preferable to optimization-based methods.Imperfect forecasts deteriorate optimization-based economic performance because cost minimization is their explicit objective.Source: §8 · p. 16
- Rule-based heuristics typically compute near-optimal schedules faster, while DP is slowest but most accurate; machine learning can use DP results for faster online solutions.This trade-off assumes a linear battery model for the heuristic speed comparison; DP uses nonlinear battery and inverter models.Source: §8 · p. 16
- When new customers have little usage data, a plug-and-play approach can assign them to a consumption-style cluster and use its representative PFA for fast online solutions.The authors describe these solutions as reasonably accurate.Source: §8 · p. 16
- Because modeling assumptions may make simulated findings differ from practical outcomes, the authors recommend caution and plan hardware experiments with real batteries.The paper identifies hardware validation as future work.Source: §8 · p. 16