Forecasting power demand has always been a critical task for energy market participants, utilities, and grid operators. But today, the stakes are higher and the challenge greater. Shifting policies around energy transition, fluctuating economic conditions, and increasingly unpredictable weather patterns – exacerbated by climate change – have introduced a new challenge to load forecasting.

As our customers reflect on their new year’s milestones and challenges, Amperon is eager to share the ways our forecasting methodologies are designed to help customers navigate complexities in demand, generation, and price forecasting.  

Amperon’s forecasting guiding principles

To design an industry-leading demand forecast, our team of data scientists, engineers, and power markets experts defined guiding principles for our solution. An industry-leading demand forecast should:  

1. Provide leading forecast accuracy which minimizes errors across every hour of the forecast

2. Deliver accurate predictions that do not overfit on historic data

3. Leverage best-in-class weather and market data

4. Provide interpretable and explainable forecasting results

5. Scale with market dynamics and customer demands by seamlessly integrating updated data, new data sets, and modeling techniques

Amperon’s demand forecasting capabilities

Amperon has developed an award-winning demand forecast serving 24 markets across North America and Europe, grounded in these guiding principles. Building on traditional approaches, Amperon brings together the best of proven regression approaches and innovative data-driven methodologies that integrate advanced machine learning (ML) techniques, robust statistical models, and efficient data pipelines. This powerful combination empowers customers to confidently meet their forecasting needs and maintain an edge in the market.

Under the hood of our dynamically weighted hybrid methodology

Predicting energy demand accurately requires a combination of modeling techniques that can identify complex patterns from weather, time-related variables (e.g. day of the week, holidays, sun-times, etc.), and historic power demand data. Starting in the 1980s, regression-based approaches became standard for short-term load forecasting due to computational advancements at that time and their strength in integrating multiple input variables. As a result, utilities quickly adopted these models to optimize power generation, grid reliability, and operational planning. Regression-based models continue to serve as a key tool in load forecasting due to their simplicity, interpretability, and effectiveness in capturing linear relationships.  

At Amperon, we began building our data science model stack with both proven linear-based regression approaches and novel techniques that machine learning has recently enabled. Our model stack can be summarized by the integration of a series of primary and secondary models that are all evaluated, calibrated, and weighted hourly to generate our final short-term demand forecast.  

PRIMARY MODELS

• Linear regression models: Regression models continue to perform well in identifying relationships between a dependent variable (electricity demand) and multiple independent variables (features from input data to our models). Specifically, we fit multidimensional curves to the observed data to capture non-linear relationships. These models are also relatively simple to explain, yet performant, providing a strong baseline for forecasting accuracy. However, relying solely on off-the-shelf linear regression models can risk inability to capture seasonal and cyclic patterns and risk high sensitivity to outliers, which can disproportionately affect the model’s performance.

• Gradient boosted tree models: We supplement regression models with boosted tree models that help our forecasts capture complex, non-linear relationships and capture potential future events that have no historic data. Through an iterative error reduction process, each tree is trained to minimize the residual errors from the previous tree by learning from data points where earlier predictions were inaccurate.  As a result, boosted tree models perform very well under sudden changes (e.g. unpredictable weather events) and can achieve low forecast error.  

SECONDARY MODELS

We utilize supplementary statistical and ML-based models to capture additional complex patterns, such as seasonality and exponential time-series, and understand the interaction of these exogenous variables and temporal patterns. By combining and dynamically weighting these diverse secondary models alongside our primary ML models, Amperon can continuously enhance its predictions through ongoing research and development of other modeling techniques.  

Amperon’s Data Science Model Stack Summary

Amperon’s forecasting advantages

Amperon has developed a distinctive and effective approach to load forecasting, guided by our five forecasting principles. By combining the strengths of ML-based models and linear regression models, this hybrid and versatile framework offers enhanced accuracy and actionable insights for our customers.

‍1. Provide leading forecast accuracy which minimizes errors across every hour of the forecast.

Amperon’s approach brings unique strengths of various modeling techniques that complement one another, leading to better performance in diverse forecasting scenarios, including extreme weather events and during demand peaks. Our models are evaluated, calibrated, and re-weighted on an hourly basis. As a result, they continuously learn and leverage insights from updated data. We optimize for every hour of the forecast with a special emphasis on peak demand.

2. Deliver accurate predictions that do not overfit on historic data.

Individual models are prone to overfitting, a phenomenon where the model mirrors training data too closely but fails to capture its underlying pattern. This is particularly true when dealing with small or noisy datasets. Overfit models are overly sensitive to minor changes in the input data and are often inaccurate when forecasting new data or novel conditions. Regression models also run the risk of underfitting when the data exhibits non-linear relationships, which ML models can capture, reducing bias in the ensemble. Our weighted model stack approach mitigates this by distributing errors and variances of individual models, resulting in a final prediction that generalizes better to unseen data.

3. Leverage best-in-class weather and market data,

Amperon combines the accuracy of weather forecasts from five industry-leading weather vendors. We apply unique ensembling techniques that analyze the performance of individual vendor forecasts and dynamically-weight inputs from multiple vendors to reduce exposure to errors from any single weather vendor. In addition, our training dataset for load, generation, and price is updated and refreshed continuously through automated data pipelines to learn from public and proprietary data sources.  

4. Provide interpretable and explainable forecasting results:  

Our forecasting approach balances the complexities of advanced feature engineering with predictive power. By combining linear regression with other machine learning models, we ensure that users can capture complex temporal patterns and residual variations while leveraging sophisticated, higher-order features. This approach offers the best of both worlds: the energy industry’s trust in regression methodologies alongside the advanced accuracy of machine learning techniques. To enhance interpretability, we provide extensive weather scenarios and robust model documentation, enabling clients to better understand the contributing factors. Additionally, our team of market and technical experts offers deep dives into market conditions and forecast results, ensuring clarity and informed decision-making.

5. Scale with market dynamics and customer demands by seamlessly integrating updated data, new data sets, and modeling techniques.

Amperon’s approach allows integration of diverse model types, allowing our capabilities to continuously improve through ongoing research and integration of new statistical, regression-based, and ML-based techniques. In addition, our framework is resilient to changing conditions and can utilize fast-learning capabilities of ML models with the consistency of linear regression.

‍

Experience the difference for yourself. Connect with us and start a trial today.  

‍