Data-driven short-term load forecast method and demand-side management for distribution network
Abstract
With the development of the power grid, the smart grid makes the system more
intelligent, efficient, sustainable and reliable with integrated Information and Communication
Systems. Moreover, the data from the advanced system provides chances to
utilise machine learning algorithms to improve the system operation further. In addition,
the load profile is undergoing altering and becoming more unpredictable because
of the increase in smart home appliances, EVs, e-heating systems, energy storage
devices, etc. These factors bring more challenges and opportunities for the future
power system to improve operational efficiency and demand response quality. In this
regard, considering load forecasting is crucial in smart distribution networks for utility
companies, especially those employing the demand-side management alternatives,
the short-term load forecast could be more accurate and robust, evolve for future load
forecasting purposes, and reveal its value in improving demand-side management
qualities.
This research proposes a novel Dynamic Adaptive Compensation-Long Short-Term
Memory (DAC-LSTM) forecast method. This method uses high time-resolution datasets
and LSTM networks as fundamental to give short-term time-series load forecast
results. The proposed method dynamically distinguishes the peak and off-peak hours
and improves the forecast accuracy separately. For DNOs, the forecast errors, especially
during peak hours, lead to penalties to start/stop backup generations or adjust
the distribution schedules, and this will result in more operational costs. Further, the
proposed method introduces a novel DAC block to compensate for forecast errors
according to the error trend calculated by historical forecast and actual load, then
applying dynamic adaptive parameters. The greater the current-to-average forecast
error ratio or the closer the forecast step to the present time stamp, the larger the
compensation factors. Besides, the factor caps are set to prevent the model from
over-compensation conditions. The sensitivity of introduced parameters is analysed,
providing the performance of the developed method under different parameter values.
Afterwards, the proposed method is evaluated with six case studies, including varying
the forecast steps (compared with LSTM and ARIMA), limiting the size and length
of the training datasets (compared with ARIMA and Persistence), comparing with
other state-of-art methods qualitatively, and comparing with ELEXON UK domestic
load forecast results. Finally, the advantages of the DAC-LSTM method are validated,
including providing accurate short-term load forecast results during peak and off-peak
simultaneously, with a shorter length of or fewer households’ historical datasets, and
compared with existing transmission network forecast methods. The system operators,
like DNOs, can reduce the operational cost with more accurate forecasts during
peak hours as well as own more load curtailment potentials during the off-peak hours.
Additionally, more contributions, including the future bottom-up load scenarios establishment
and the improved Stackelberg Game demand response for end-user utility
bill reductions, will help system operators develop suitable DSM alternatives and
tariffs based on more realistic and accurate analysis. To be more specific, first, based
on the Ten Year Network Development Plan (TYNDP) 2018 and the UK government
reports, bottom-up load profiles are designed and generated for the UK distribution
network for the scenario years 2020, 2030 and 2040. The DAC-LSTM method is
evaluated with these scenario profiles, yielding up to 0.989 and 3.79% (measured in
R2 and MAPE) forecast accuracy, for various levels of electric vehicle and e-heating
penetration when compared with the ARIMA and Persistence methods. Second, a
DSM alternative is built based on a Stackelberg Game to reduce the consumers’ utility
bill, which considered the forecast error as a constraint. In this game, given that the
consumer offers maximum controllable power to the operator, the game achieves the
Stackelberg Equilibrium while maximising the operator’s revenue and supplying the
necessary power to consumers. The case study demonstrates that when considering
forecast errors in demand response strategies, higher forecast accuracies reduce the
electricity bill up to 10.4% in an ideal circumstance. The improved Stackelberg Game
makes the forecast error one primary constraint that most existing DSM alternatives
lack. This proves the value of utilising state-of-art forecast methods in the deployment
of DSM alternatives.