Source-linked AI summary
Digital Twins based Day-ahead Integrated Energy System Scheduling under Load and Renewable Energy Uncertainties
Minglei You, Qian Wang, Hongjian Sun, Ivan Castro, Jing Jiang
TL;DR
Practical integrated energy systems face multiple uncertainty sources that complicate day-ahead scheduling. This paper uses a digital twin with a virtual multi-vector energy-system replica and a deep neural network trained on forecasts and historical errors, reducing uncertainty-related operating costs by 63.5% versus a forecast-based benchmark.
Problem
Multiple uncertainty sources limit the practicality of day-ahead IES scheduling, while prior work often addresses single sources or relies on forecasting assumptions.
Method
A digital-twin IES uses a virtual MVES replica and a deep-learning scheduler trained from historical forecasting errors and day-ahead forecasts.
Results
63.5%: the proposed method reduces average daily extra operating cost induced by multiple uncertainties versus the benchmark method.
Takeaways & Limitations
The results identify EVs and TESs as proactive components in the proposed scheduling method for future energy-system integration and decarbonisation.
Abstract
from arXiv · showhide
By constructing digital twins (DT) of an integrated energy system (IES), one can benefit from DT's predictive capabilities to improve coordinations among various energy converters, hence enhancing energy efficiency, cost savings and carbon emission reduction. This paper is motivated by the fact that practical IESs suffer from multiple uncertainty sources, and complicated surrounding environment. To address this problem, a novel DT-based day-ahead scheduling method is proposed. The physical IES is modelled as a multi-vector energy system in its virtual space that interacts with the physical IES to manipulate its operations. A deep neural network is trained to make statistical cost-saving scheduling by learning from both historical forecasting errors and day-ahead forecasts. Case studies of IESs show that the proposed DT-based method is able to reduce the operating cost of IES by 63.5%, comparing to the existing forecast-based scheduling methods. It is also found that both electric vehicles and thermal energy storages play proactive roles in the proposed method, highlighting their importance in future energy system integration and decarbonisation.
1. Introduction
Integrated energy systems can improve efficiency and decarbonization through multi-vector coordination, but practical scheduling must handle multiple, potentially unknown uncertainty sources. The paper therefore proposes a digital-twin and deep-learning method that directly produces day-ahead schedules for uncertain IES operation.
- Digital-twin motivation: Digital twins link a virtual MVES replica with the physical IES and external environment so system data can support simulation, optimization, prediction, and operational coordination.Predictions or optimized configurations from the virtual replica can be sent to physical devices, potentially improving coordination and reducing operating costs.
- Research gap: The paper targets day-ahead IES scheduling when multiple uncertainty sources coexist, extending beyond prior work focused on isolated uncertainty sources or forecasting accuracy.Renewable intermittency and forecasting errors are identified as practical uncertainty sources, while prior methods often assume known distributions or address single sources.
- Proposed contribution: The proposed method addresses multiple uncertainties by directly generating day-ahead scheduling values rather than using machine learning only to improve forecasts.This distinguishes the scheduling objective from related approaches that integrate forecasts into dispatch or scheduling.
- Proposed contribution: A data-augmentation-based deep-learning method learns cost-saving schedules from historical forecasting errors and day-ahead forecasts.The contribution is presented as a practical approach for real-life IESs containing multiple uncertainty sources.
- Evaluation: The study evaluates the proposed approach with real-world U.K. data and reports its conclusions after case studies of IES scheduling.The paper states that case studies are performed using real-world U.K. data.
2. System Model
The system model represents an IES as an aggregated multi-vector network containing electrical, thermal, renewable, conversion, storage, and vehicle components. Its MVES virtual replica exchanges information with the physical system, makes day-ahead device schedules, and coordinates real-time operation under physical constraints.
- Multi-vector Energy System Model: The aggregated IES supplies electricity and heat using electricity, renewable energy, and natural gas through transformers, boilers, CHP systems, TESs, and EVs.Aggregation allows component classes such as wind farms or large building loads to be represented as grouped system elements.
- Digital-twin integration: The MVES virtual replica coordinates physical-device data, uses forecasts, market prices, and storage status for day-ahead decisions, then monitors and coordinates real-time operation.Day-ahead plans include TES and EV decisions and are stored for next-day operations.
- Energy-flow representation: The model includes imported electricity and natural gas, wind turbines, solar panels, EVs, and TES systems with energy flows defined for each scheduling slot.Energy-flow signs distinguish energy injected into the IES from energy sourced from it.
- Operational constraints: Physical constraints require electricity supplied by the IES to meet electricity-load demand at each scheduling slot.The electricity-load constraint is expressed through the system energy-balance equation and converter efficiencies.
- Operational constraints: The model also imposes heat-load and converter-flow constraints, including maximum energy-flow limits and efficiency parameters for CHP systems and boilers.The supplied formulation identifies CHP thermal recovery efficiency and boiler energy efficiency among the relevant parameters.
W , SMAX
The scheduling formulation models EVs and TESs with operational and state constraints, then minimizes expected total cost under forecast errors using day-ahead forecasts and real-time adjustments.
- EV and TES models constrain charging, discharging, state of charge, and restoration of their original energy levels after service.The EV and TES formulations include bounded operating states and zero net energy flow over their service periods.
- Real-time mismatches between actual and forecast electricity, heat, wind, and solar quantities are compensated through electricity and natural-gas market transactions.EV and TES charging and discharging continue to follow their day-ahead schedules during real-time operation.
- The objective minimizes expected actual total cost over the 24-hour horizon while accounting for potential forecasting errors.The expectation uses the probability distribution of the uncertain variables, and scheduling is based on day-ahead forecasts.
- 63.5% lower operating cost is reported for the proposed DT-based method versus existing forecast-based scheduling methods.The supplied paper context reports this overall case-study comparison.
3. Deep Learning Embedded Integrated Energy System Scheduling Scheme
The scheme trains a deep neural network to produce statistical day-ahead schedules from forecasts and sampled forecasting errors, while enforcing physical constraints through output transformations and loss penalties.
- The proposed scheme learns statistical optimal day-ahead scheduling from historical forecasts and forecasting errors for operation under multiple uncertainties.
- The DNN maps day-ahead forecasts Ft to scheduling outputs St, and the trained network can commit schedules for different inputs without a new optimization search.
- Physical constraints are handled through output-layer constraint enforcement, scaled energy-flow outputs, daily-average adjustments, and penalty terms for storage state of charge.
- Forecasting errors sampled from historical observations or simulations augment each forecast into potential actual-value datasets for statistical training.
- The loss evaluates schedules across potential actual values, while gradient descent updates the DNN parameters during training.
4. Case Studies
Case studies evaluate the DT-based scheduler across hourly, daily, cost-category, cross-year, constraint-adjustment, and larger-system settings. The method generally reduces costs versus forecast-direct scheduling while actively using EVs and TESs under uncertainty.
- 4.1. A Case Study on the Hourly Cost Performance: The proposed method reduces costs for most hours, with exceptions at the 3rd, 7th, 20th, and 23rd hours.Its average execution time is 0.002 s, compared with 0.355 s for the benchmark.
- 4.2. A Case Study on the Daily Cost Performance: 27 of 31 days show reduced daily costs with the proposed method in May 2019.The result reflects day-level coordination of IES devices over complete 24-hour scheduling periods.
- 4.3. A Case Study on the EV and TES Performance: EV and TES payments increase by 18.2% and 56%, respectively, as the scheduler uses them more actively across the day.The increased energy-buffer usage accompanies reduced extra costs and lower daily costs on most days.
- 4.4. A Case Study on the 2018 U.K. Dataset: The trained model retains cost-reduction performance on 2018 data after training on 2019 data, indicating a practical statistical scheduling solution.The cross-year evaluation tests the model on an unknown dataset without theoretical optimal schedules.
- 4.4. A Case Study on the 2018 U.K. Dataset: 63.5% lower average daily extra operating cost is achieved than forecast-based scheduling on the 2018 U.K. dataset.The benchmark incurs £34.5 in daily extra cost, versus £12.6 for the proposed method.
- 4.5. A Case Study on Physical Constraint Adjustment Costs: Physical constraint adjustment adds £0.0012 on average, a marginal amount relative to daily costs.Constraint enforcement ensures practical feasibility, although adjustment can incur additional cost.
5. Conclusions
The paper proposes a DT-based day-ahead scheduling scheme whose deep network learns from forecasts and historical errors while enforcing physical constraints. Across U.K. case studies, it reduces extra operating cost and actively schedules EVs and TESs, while future work targets broader carbon, forecasting, system-size, and real-time settings.
- 5. Conclusions: The proposed DT-based scheme combines virtual–physical IES interaction with deep learning from historical forecasting errors and day-ahead forecasts.Data augmentation addresses multiple uncertainties, while network and loss designs enforce physical constraints.
- 5. Conclusions: The method reduces average daily extra operating cost by 63.5% compared with benchmark scheduling that uses forecasts directly.The evaluation uses historical U.K. data from 2018 and 2019.
- 5. Conclusions: Active EV and TES scheduling is presented as a way to address multiple uncertainty challenges in future energy systems.The conclusion links this role to progress toward the net-zero target.
- 5. Conclusions: Future work will examine carbon-emission reduction, rolling intraday forecasts, larger systems, advanced real-time scheduling, and alternative data augmentation methods.These extensions are proposed to broaden evaluation and potentially reduce operating costs further.