💻 computer science

Seasonality-Aware Dual-Memory Pricing for Electric Vehicle Charging Stations under Non-Stationary Demand

This paper introduces Seasonality-Aware Dual-Memory Pricing (SA-DMP), a learning algorithm that dynamically balances long-term seasonal patterns and short-term market shifts to optimize electric vehicle charging prices under non-stationary demand, demonstrating statistically significant regret reduction compared to fixed-weight baselines while identifying specific operational boundaries where adaptive pricing outperforms traditional time-of-use tariffs.

Original author: Xinzi Wang

Published 2026-11-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original author: Xinzi Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

Electric vehicle charging stations face a constant balancing act. Operators must set prices that attract drivers while covering the cost of electricity, all without knowing exactly how many people will arrive or how much they are willing to pay at any given moment. This is a problem of learning from limited clues. When a station posts a price, it only sees the result of that specific choice; it never learns what would have happened if it had chosen a different price. This creates a puzzle where the rules of the game change over time. Demand follows predictable patterns, like rush hour peaks or quiet weekends, but it also shifts unexpectedly when a new competitor opens nearby or when local habits change. The challenge for computer systems managing these stations is to remember the reliable, repeating patterns of the past while staying flexible enough to adapt to the sudden changes of the present.

Researchers at Shenyang University have developed a new method to solve this specific puzzle, called seasonality-aware dual-memory pricing. Instead of relying on a single way of looking at the past, their system maintains two separate memories of demand. One memory looks far back, capturing the steady, repeating rhythms of the week and the year. The other memory looks only at the very recent past, focusing on the immediate conditions of the market. The innovation lies in how the system decides which memory to trust. Rather than using a fixed rule to blend these two views, the system constantly watches its own performance. If the recent memory has been making better predictions lately, the system leans more heavily on it. If the long-term patterns prove more reliable, it shifts its focus back to the broader history. This adjustment happens automatically, based on the errors the system made in previous hours.

To test this idea, the researchers built a detailed computer simulation of a charging station. They created six different worlds, each with its own unique demand behavior. In some scenarios, the demand was steady and predictable. In others, it changed abruptly, drifted slowly over time, or was hit by temporary shocks that vanished as quickly as they appeared. They ran their new system against a version that used a fixed, unchanging blend of the two memories, as well as against several other standard methods used in the field. The results showed that the new, adaptive system consistently made better pricing decisions. In every single scenario tested, it reduced the difference between its performance and the theoretical best possible outcome. On average, this improvement translated to a reduction in lost potential value of about 0.44 percentage points. While that number might seem small, in the context of continuous pricing over long periods, it represents a meaningful gain in efficiency.

The study also looked at what happens when the system is tested in a more realistic, complex environment where charging sessions can last for several hours and occupy physical plugs. Here, the researchers compared the adaptive system not just to a fixed blend, but also to a carefully selected standard tariff schedule. The adaptive system did outperform the fixed blend, proving that the ability to shift between memories adds value even when the physical constraints of the station are taken into account. However, the adaptive system did not beat the carefully selected standard tariff in terms of total profit. This finding is crucial: it suggests that while the new method is excellent at fine-tuning how the system learns from the past, it does not yet replace the need for a well-designed, fixed pricing schedule. The adaptive gate is a powerful tool for learning, but it works best when it complements a solid foundation rather than trying to build the entire strategy from scratch.

The researchers found that the system's success depended heavily on the type of change it faced. It performed exceptionally well when demand was hit by temporary shocks, such as a sudden, brief drop in customers that quickly returned to normal. In these cases, the ability to rapidly trust the recent past and then quickly return to trusting the long-term history was a clear advantage. Conversely, when the demand changed in a slow, steady drift, the benefit of the adaptive system was smaller, as a fixed blend of memories worked almost as well. The system also showed that it could be confused by irregular, unpredictable changes, sometimes overreacting to temporary noise. This tells us that the method is not a magic bullet for every situation, but a specialized tool that shines when the environment oscillates between stability and sudden, short-lived disruption.

Ultimately, this work demonstrates that for electric vehicle charging, the way a computer learns from its mistakes matters as much as the data it sees. By giving the system the ability to weigh its own recent performance against its long-term experience, the researchers created a pricing strategy that is both stable and responsive. The findings suggest that future charging networks could become more efficient by using these adaptive learning rules to manage prices in real-time. However, the study also draws a clear boundary: this learning mechanism improves the fine-tuning of decisions, but it does not yet surpass the value of a well-calibrated, fixed pricing plan in generating maximum profit. The path forward involves combining this smart, adaptive learning with the operational realities of physical charging stations, ensuring that the system learns not just from numbers, but from the actual flow of cars and energy.

Technical Summary: Seasonality-Aware Dual-Memory Pricing for Electric Vehicle Charging Stations under Non-Stationary Demand

Problem Formulation
Electric vehicle (EV) charging operators face a dynamic pricing challenge characterized by non-stationary demand that combines recurring seasonal patterns (e.g., daily and weekly cycles) with structural changes (e.g., shifts in customer composition or market entry). The decision process is a multi-armed bandit problem with partial feedback: operators select a price from a discrete set and observe only the demand resulting from that specific price, not the counterfactual demand at other prices. The core difficulty lies in balancing the retention of reusable seasonal information against the need to rapidly adapt to recent market conditions. Existing non-stationary bandit methods typically rely on fixed forgetting rules (e.g., sliding windows or discounting) that do not dynamically allocate weight between long-term seasonal memory and short-term market evidence based on current prediction performance.

Methodology: SA-DMP
The author proposes Seasonality-Aware Dual-Memory Pricing (SA-DMP), an online learning algorithm designed to address this tension. The method operates through three primary components:

Dual-Memory Forecasting: The system maintains two demand predictors for each price arm and calendar context:
- A long-memory estimator ( $H^L = 672$ hours) that captures recurring seasonal patterns using ridge-regularized harmonic features.
- A short-memory estimator ( $H^S = 168$ hours) that emphasizes recent market states.
- Both estimators are seasonally adjusted using harmonic features to account for weekly variations without mechanically equating demand across different prices.
Loss-Adaptive Blending Gate: Instead of using a fixed mixture weight, SA-DMP employs an online gate that dynamically adjusts the weight ( $\rho$ ) assigned to the short-memory predictor based on previous prediction losses.
- The algorithm tracks a recursively discounted statistic ( $G_{c,t}$ ) representing the squared-error advantage of the short model over the long model.
- A logistic function converts this advantage statistic into a mixing weight $\rho_{c,t}$ . If the short model accumulates lower error recently, its weight increases; if the long model performs better, the weight decreases.
- This allows the system to adaptively trust recent observations during structural shifts while reverting to seasonal history when recent data is noisy or transient.
Optimistic Price Selection: The blended demand estimate is used to construct an optimistic served-demand score incorporating a count-based exploration bonus (UCB-style). The operator selects the price that maximizes the expected contribution margin (price minus procurement cost) multiplied by this optimistic score.

Key Contributions
The paper isolates the specific contribution of adaptive memory weighting through a rigorous, architecture-matched experimental design:

Benchmark Formulation: A reproducible EV charging pricing benchmark using selected-price feedback and a separately computed dynamic pseudo-regret comparator.
Mechanistic Isolation: By matching the seasonal estimators, observation histories, and price-selection rules between SA-DMP and a fixed-weight dual-memory baseline (Fixed- $\rho = 0.8$ ), the study isolates the incremental value of the online loss-adaptive gate.
Operational Evaluation: A complementary simulation evaluates the method under multi-hour connector occupancy and abandonment constraints, distinguishing learning performance (regret) from realized operational outcomes (contribution margin).

Results
The evaluation was conducted across six controlled demand environments (Stationary, Abrupt, Gradual, Periodic, Mixed, and Seasonal Shocks) over 100-day horizons with 20 paired seeds per scenario.

Regret Reduction: SA-DMP achieved lower mean cumulative dynamic pseudo-regret than the fixed-weight baseline in all six environments. The macro-average oracle-normalized reduction was 0.441 percentage points (95% bootstrap interval: 0.315–0.564).
Scenario Sensitivity: The gains were most pronounced in the "Seasonal Shocks" scenario (0.760 pp), where the ability to rapidly switch between trusting recent data and long history is critical. Gains were smallest in "Gradual" drift (0.276 pp), where a fixed blend remains near-optimal.
Calibration Robustness: In a post-development study using disjoint test seeds, SA-DMP outperformed a validation-selected fixed blend ( $\rho=0.85$ ) with a macro-average reduction of 0.463 percentage points.
Distribution Shifts: Under specific distribution shifts (e.g., irregular regime changes), the adaptive gate occasionally underperformed the fixed blend, indicating that the advantage is contingent on the structure of the demand variation.
Operational Margin: In the eight-connector operational simulation, SA-DMP yielded higher mean contribution margins than the matched fixed blend in all three tested scenarios. However, absolute margins remained below those achieved by a validation-selected Time-of-Use (TOU) tariff, delineating the boundary where adaptive learning complements rather than replaces calibrated static tariffs.

Significance and Claims
The paper claims that the primary value of SA-DMP lies in the incremental improvement of learning efficiency under non-stationary, partial-feedback conditions. By dynamically weighting seasonal memory against recent evidence based on prediction loss, the method avoids the pitfalls of both excessive forgetting (discarding useful seasonal patterns) and excessive inertia (failing to adapt to structural changes).

The author emphasizes that this approach identifies the specific value of loss-adaptive memory weighting distinct from general policy learning or tariff selection. While the adaptive gate improves performance over fixed-weight baselines, the study modestly concludes that in the tested operational environments, it serves as a complement to, rather than a replacement for, well-calibrated static or TOU tariffs. The findings suggest that the method is most effective when demand variations involve temporary shocks or cycles where the divergence between short- and long-term predictors is significant and self-correcting.

More like this