Utilizing Device-level Demand Forecasting for Flexibility Markets - Full Version (1805.00702v1)

Abstract: The uncertainty in the power supply due to fluctuating Renewable Energy Sources (RES) has severe (financial and other) implications for energy market players. In this paper, we present a device-level Demand Response (DR) scheme that captures the atomic (all available) flexibilities in energy demand and provides the largest possible solution space to generate demand/supply schedules that minimize market imbalances. We evaluate the effectiveness and feasibility of widely used forecasting models for device-level flexibility analysis. In a typical device-level flexibility forecast, a market player is more concerned with the \textit{utility} that the demand flexibility brings to the market, rather than the intrinsic forecast accuracy. In this regard, we provide comprehensive predictive modeling and scheduling of demand flexibility from household appliances to demonstrate the (financial and otherwise) viability of introducing flexibility-based DR in the Danish/Nordic market. Further, we investigate the correlation between the potential utility and the accuracy of the demand forecast model. Furthermore, we perform a number of experiments to determine the data granularity that provides the best financial reward to market players for adopting the proposed DR scheme. A cost-benefit analysis of forecast results shows that even with somewhat low forecast accuracy, market players can achieve regulation cost savings of 54% of the theoretically optimal.

• The paper demonstrates that device-level demand forecasting effectively manages renewable energy uncertainty via detailed flex-offer aggregation.
• It introduces a two-step approach combining Logistic Regression and Pattern Sequence Matching to predict device activation and estimate energy profiles.
• The study finds that aggregating forecasts to a group level offers the best trade-off, achieving significant regulation cost savings despite forecast errors.

This paper explores the practical application and financial viability of using device-level demand forecasting for flexibility markets, specifically focusing on the Nordic energy market context with high Renewable Energy Source (RES) penetration (1805.00702). The core idea is to manage the uncertainty introduced by fluctuating RES by leveraging the inherent flexibility in energy consumption of individual household appliances.

1. Demand Flexibility and Flex-Offers

• Concept: Demand flexibility is the ability to shift energy consumption in time (time flexibility) or adjust the amount consumed (amount flexibility) within certain constraints. The paper focuses on device-level flexibility (e.g., washing machines, dishwashers) because it provides the most detailed ("atomic") information, leading to the largest possible solution space for optimization compared to aggregated household-level data.
• Representation: Flexibility is modeled using "flex-offers". A flex-offer for a device contains:
  • An energy profile: Expected energy consumption per time unit (e.g., 15 minutes) during operation.
  • A time flexibility interval: The earliest and latest possible start times for the device's operation.
• Lifecycle: Flex-offers follow a lifecycle:
  1. Generation: Forecast future device usage and associated flexibility based on historical data and patterns.
  2. Aggregation: Combine small, device-level (micro) flex-offers into larger (macro) flex-offers suitable for market trading by an Aggregator entity.
  3. Scheduling/Trading: Market players (like Balance Responsible Parties - BRPs) schedule or trade these macro flex-offers to optimize their energy balance, often to minimize costs associated with the regulating power market.
  4. Disaggregation: Translate the scheduled macro flex-offers back into specific operating times for individual devices.
  5. Execution: Smart devices automatically adjust their operation based on the received schedules. Users typically retain override capability but may lose incentives.

2. Forecasting Methodology

The paper proposes a two-step approach for forecasting device-level demand, acknowledging the intermittent nature of appliances like washer dryers:

• Step 1: Activation Prediction: Uses Logistic Regression (LR) to predict the probability of a device activating (switching on from an idle state) within a specific time interval (hourly, group of hours, or daily).

• Step 2: Demand and Duration Estimation: If activation is predicted, Pattern Sequence Matching (PSM) is used to estimate the likely duration of operation and the energy consumption profile for that operation. PSM finds historical operations that started at similar times (or days) and averages their profiles and durations.

  • PSM Algorithm (Hourly Example):

  1. Input: Historical time series data ($X$), predicted activation hour ($h$). 2. Find all past instances where the device activated at hour $h$. 3. Extract the energy profile ($p_i$) and duration ($|p_i|$) for each historical activation. 4. Calculate the average duration ($l$) across these historical instances (using ceiling). 5. Calculate the average energy consumption for each time step ($j$) up to duration $l$ across the historical profiles. 6. Output: The estimated energy profile $p$ for the predicted activation. (Similar logic applies to group and daily resolutions, adjusting the search criteria in PSM).

• Data and Resolution:

  • Uses 15-minute device-level power readings from smart plugs (Jan 2014 - Oct 2015) and hourly Danish energy market data (spot/regulation prices/volumes).
  • Experiments with three data granularities for forecasting:
  • Hourly: Predict activation for each of the next 24 hours. (Binary vector $a_i \in \{0,1\}^{24}$)
  • Group: Cluster hours (e.g., morning, afternoon, night) and predict activation within each group. (Binary vector $a_i \in \{0,1\}^m$)
  • Daily: Predict if the device will activate at least once during the next day. (Binary value $a_i \in \{0,1\}$)
• Feature Engineering: To improve LR prediction accuracy, several features are extracted:
  • Last 24 hourly states (or last 7 daily states for daily prediction).
  • Hour of the day (one-hot encoded).
  • Day of the week (one-hot encoded).
  • Weekend flag.
  • Time since last operation (binned, one-hot encoded).
  • Season (one-hot encoded).
  • Interaction features (multiplicative combinations).
• Regularization: L1 regularization is used in the LR model to prevent overfitting due to the high number of features and automatically perform feature selection.
• Evaluation Metrics: Due to the highly imbalanced nature of device activation data (mostly 'off'), Area Under the Precision-Recall curve (AUC-PR) and F1-score are used instead of ROC AUC. Forecast outcomes are categorized as True Positives (TP), False Positives (FP), False Negatives (FN), and True Negatives (TN).

3. Financial Evaluation

The paper quantifies the financial benefit (savings in the regulation market) and loss (due to forecast errors) associated with using the forecasted flexibility.

• Regulation Market Context: BRPs face costs when their actual supply/demand deviates from their spot market commitments, forcing them to buy/sell power in the more expensive regulating market. Flexibility allows BRPs to shift demand to counteract these imbalances.
• Scheduling Optimization: An optimization problem is formulated to schedule the forecasted flexible demands ($\hat{F}$) within their time flexibility ($\tau$) to maximize the reduction in regulation volume ($V$). This is solved using the GLPK solver via PuLP. The computational complexity is noted as low for typical device-level scenarios.
• Quantifying Savings and Loss:
  • Potential Savings ($\Delta R$): The reduction in regulation cost achieved by optimally scheduling the forecasted flexible demand, compared to the baseline cost without scheduling. This primarily benefits TPs.
  • Loss ($L$): The cost incurred due to forecast errors. FPs lead to scheduling phantom demand, potentially worsening imbalances. FNs mean unexpected demand appears, also potentially worsening imbalances. The loss calculation considers both the volume difference ($|f(i) - \hat{f}(i)|$) and the impact of the error on regulation prices (using a price-volume relationship model, Eq. 3).
  • Net Benefit: $\Delta R - L$.

4. Experimental Results and Practical Implications

Experiments were conducted using real washer dryer data and Danish market data.

• Forecast Accuracy vs. Granularity:
  • Accuracy (AUC-PR) significantly improves with data aggregation: Daily (0.84) > Group (improved AUC) > Hourly (0.23).
  • Hourly prediction is challenging due to high stochasticity. Group resolution captures broader patterns (e.g., usage in evenings). Daily prediction captures the general propensity to use the device.
• Financial Savings vs. Accuracy:
  • Key Finding: Substantial regulation cost savings (up to 54% of theoretical maximum) are achievable even with relatively low forecast accuracy at finer granularities (e.g., 42% savings with hourly AUC-PR of 0.23, corresponding to Precision/Recall around 0.3).
  • The financial utility is not directly proportional to standard accuracy metrics like AUC-PR. Even many FPs don't necessarily negate savings, as their impact depends on the market conditions at the time of the error.
  • Group resolution provides the best trade-off: It yields the highest savings (54% optimal) by balancing improved forecast accuracy over hourly with greater retained flexibility compared to daily.
  • Daily resolution: Highest accuracy but lowest savings due to the significant loss of flexibility (only knowing if it runs, not when within the day for granular scheduling).
  • Hourly resolution: Lowest accuracy but still achieves significant savings (42% optimal), demonstrating viability even with high uncertainty.
• Time Flexibility: Savings increase as the allowed time flexibility ($\tau$) increases, giving the scheduler more options.
• Threshold Selection: The paper proposes a practical rule of thumb: selecting the probability threshold for the LR classifier that yields the highest F1-score tends to result in near-optimal financial savings. This helps bridge the gap between model tuning and achieving the desired economic outcome.

5. Conclusion

The paper demonstrates the financial feasibility of incorporating device-level demand flexibility into energy markets. Even with imperfect forecasts typical of stochastic device usage, significant savings in regulation costs can be achieved. Aggregating data to a "group" level (clusters of hours) offers the best balance between forecast accuracy and usable flexibility, maximizing financial benefits. The F1-score provides a practical heuristic for tuning forecast models (specifically, setting the decision threshold) to optimize for market savings. This approach offers a promising way for market players like BRPs to mitigate imbalance costs associated with RES integration.