Leading-Edge Electricity AI Forecasting Features

Leverage our cutting-edge, no-code Auto-ML platform featuring automated feature engineering, AI modeling, hyperparameter optimization and backtesting, enabling users to build, test, and deploy robust, high-accuracy forecasting solutions in record time.

QR AI Forecaster Offers Advanced Features for Electricity Forecasting

QR AI Forecaster is a universal AI energy forecasting platform providing extensive advanced features and capabilities designed to address the full range of demand, price and renewables forecasting requirements.

Forecast Temporal Resolution & Frequency

Our standard forecast timescale and frequency for price, load & demand, and renewable generation:

Data	Forecast Resolution	Forecast Points per Execution	Execution Update Frequency	Forecast Horizon	Forecast Points per Day	Probabilistic Percentiles
RDT 5-min	5-min	288	Every 5-min 288 Forecast / Day	24 h	288 x 288	5,10,50,90,95
RTD 15-min	15-min	96	Every 15-min 96 Forecasts / day	24 h	96 x 96	5,10,50,90,95
RTD hourly*	Hourly	24	Hourly 24 Forecasts / day	24 h	24 x 24	5,10,50,90,95
DA x 14 **	Hourly	24 x 14	6, 7, 8 am	14 DA	3 x 14 x 24	1-99
DAM ***	hr	24	6, 7, 8 am	24 h	3 x 24	1-99

* Lower-timescale intraday RTD price or load, e.g., 5-min, are scaled to hourly data and forecasted in real-time.

** DA stands for any day-ahead forecast such as price, load or renewables.

*** DAM refers to the day-ahead electricity market price which settles once a day after markets close.

Multi-resolution AI forecasting:

Forecasts are generated at the native temporal resolution of the data (5–60 minutes) and can also be produced at coarser time resolutions through temporal aggregation. For example, 5-minute time series can generate 5, 15, 30, and 60-minute forecasts.
The forecasting workflow follows a fine-to-coarse modeling approach. For example high-resolution data such as 5-minute meter readings may contain artificial volatility due to metering or communication noise. To reduce overfitting, the AI engine can aggregate the data to a lower resolution (e.g., 30 minutes) for training and forecasting, and then disaggregate the results back to the original 5-minute granularity.

Feature Engineering

No-coding feature engineering

QR AI Forecaster provides an advanced, no-code, modular, and composable feature-engineering platform delivered through a dedicated web-based dashboard. The platform enables users to visually construct complex, deterministic preprocessing pipelines by drag-and-drop composition of parameterized mathematical processors. These pipelines are applied uniformly to both target time series (e.g., load, price, renewable generation) and associated predictor variables, ensuring consistent, reproducible feature transformations across training, back-testing, and production forecasting workflows.

Automated feature selection

Contextual and data engineering features are discussed elsewhere. In addition, users can add unlimited number of external time series to the feature set of an AI model as desired. For example, weather data to AI demand models, irradiance to Solar generation AI models, and for AI price models, outages, system demand, or solar and wind generation. Furthermore, the visual interface allows users to implement feature as functions of other features, e.g., demand / solar generation. All time series features are ranked for relevance to the target time series to be forecasted by different algorithms and the best are automatically linked to the AI model.

No-coding on-screen drag & drop feature engineering adding time series

On-screen best feature selection processing using different algorithms.

No-coding on-screen drag & drop feature engineering: weather, radiance and other time series.

Data Processing Engineering

AI forecasting in electricity relies on large volumes of training data, much of which is ingested in real time. Such data contains missing values, erroneous and out-of-bounds measurements, and poorly distributed historical samples. At the point of ingestion by our datahub, data is stored exactly as received, to preserve a faithful and auditable raw data record. This ensures data lineage, traceability, and regulatory compliance. However, using raw data with such deficiencies for AI model training results in severely degraded forecast accuracy.

Data challenges are addressed by our automated mathematical data preprocessing pipeline. This no-coding pipeline allows users to drag and drop modular mathematical processors and attach them to individual time series to clean, transform, and enhance the data prior to forecasting.

On-screen load data preprocessing (gaps, outliers, quantization), contextual feature engineering

Bottom up load forecasting model: aggregating load pre-forecasting and making a single forecast.

Outlier detection and replacement are critical for AI modeling, as out-of-bound values mislead AI training by reinforcing exceptional behaviors that should not be learned or reproduced. Methods available are statistical and filtering.

Gap filling is fundamental for AI forecasting, as missing data cause training failure. Methods provided are (non-) linear interpolation, machine learning, statistical and profile based.

Quantization organizes scattered data into a few classes easier for the AI to interpret. For example, outages. Machine Learning,

Smoothing and noise reduction methods can be applied to mitigate noise in data. Various moving average and filtering methods with configurable window size are supported.

Scaling or normalization ensure numerical stability and balanced feature influence, especially when combining heterogeneous inputs. The platform supports multiple.

Data augmentation adds robustness to AI models by stretching the value range of historic data used for training to allow the AI to forecast current higher or lower levels the AI would not have seen during training otherwise.

Contextual Feature Engineering

Calendar and holiday behavioral normalization

Two important categories of features engineering namely data processing and contextual feature engineering were discusses. Public holidays exhibit consumption and pricing dynamics that differ significantly from standard weekday patterns. To ensure behavioral consistency and improve forecast stability, we apply a rule-based temporal normalization framework. The AI engine systematically maps mid-week holiday data to the nearest weekend profile (Saturday or Sunday) both in training and forecasting. This maintains behavioral continuity in autoregressive features and prevents distortion from irregular holiday effects, while aligning historical learning and forward looking prediction logic. Calendar features significantly enhance predictive stability. Time-awareness contextualization These features embed structured temporal information directly into the AI modeling framework: month and week of year, day of week, hour of day. Unlike purely autoregressive features, temporal dimensions enable the AI model to learn seasonal patterns, weekly and intraday behavioral, and predict structural shifts in the absence of recent lag signals e.g., after holidays or regime changes.

Peer-point contextual features

Lagged indicators are highly valuable features in time-series AI models. They enable the AI model to capture temporal dependencies and recurring behavioral patterns embedded in historical data. Some examples are:

Previous time-interval values, e.g., last 5, 15 or 60 min.
Same-hour previous day values.
Same interval value from previous day or week
Rolling averages and lagged volatility measures.

By incorporating these structured lags, the model can learn short-term persistent effects (autocorrelation) and patterns of behavior or regime-dependent temporal correlations. These significantly enhance predictive stability

No-coding on-screen contextual feature engineering: similar points, and temporal features.

AI No-coding auto-ML AI Platform

QR AI Forecaster is a no-coding auto-ML platform

QR AI Forecaster is a no-coding AutoML forecasting platform designed for advanced AI analytical workflows. It provides a dedicated web-based dashboard in which all components of an AI forecasting job are visually configured using drag-and-drop of components and processors. AI governance Each AI forecast job is fully sandboxed and uniquely identified by a job ID. All model configurations, parameters and metadata are persistently stored and transparently exposed through the job web-dashboard, enabling full control, reproducibility and auditability. AI Model configuration Using a drag-and-drop visual interface, machine learning and deep learning model components can be assembled, including hyperparameters, training window, validation strategy, forecast horizon, execution parameters, and probabilistic model settings, along with data preprocessing and feature engineering components described elsewhere. Auto-ML and AI model optimization. The AI engine is self-calibrating and automatically determines the optimal set of model hyperparameters for the given dataset by minimizing training error. Hyperparameter optimization is performed using grid search and genetic optimization methods.

Profile & pattern making

Electricity data have clear seasonality patterns, e.g., daily off/peak, or sunlight profiles, weekdays and weekends, summer and winer. QR AI machine allows defining splitters in custom blocks or profiles. Each block is automatically assigned to its own AI model, and its training data is obtained by splitting the main time series at run-time according to the selected profiles. For example, by applying an off/peak splitter, the AI model actually splits and trains 2 different AI models.

Probabilistic forecast

To explicitly quantify forecast uncertainty and risk, the platform supports probabilistic forecasting as a native capability. When enabled, the model outputs predictive distributions expressed as 1 to 99 percentile forecasts. Typically the 5th percentile represents a lower bound, and the 95th an upper bound.

Training and forecast windows

Each forecast job defines explicitly the start and end dates for both training and forecasting periods. Seasonal regime changes require periodic adjustment of the training window to prevent contamination from mixed seasonal patterns. As the season progresses, the training interval is dynamically updated to maintain model relevance and accuracy.

Real-time & backtest mode

Multiple execution modes are supported by the AI architecture:
Production Mode is used for forward-looking forecasting, including automated data ingestion, scheduled model execution, operational deployment outputs
Back-test Mode is specifically designed to perform rigorous electricity back-test forecasting, replicating real production conditions by sequentially forecasting each period within the historical test window, while freezing the dataset at each forecast point so as to only use the data available prior to the forecast timestamp, strictly excludes any future information beyond the forecast point. This rolling, step-by-step simulation ensures that forecasts are generated exactly as they would have been in production mode, without look-ahead bias.

Forecast accuracy

Forecast accuracy is measured using MAPE and MAE, with additional accuracy metrics easily configurable as needed. The forecast dashboard presents actual versus forecasted values in both tabular and graphical formats, along with period-level MAPE and MAE and overall MAPE and MAE for the entire forecast horizon. The forecast dashboard continuously monitors predefined MAPE threshold levels and displays visual error indicators, including red warning gauges when thresholds are exceeded. Poor forecast performance automatically triggers alerts, prompting the team to review and adjust model features, parameters and assumptions.

Bottom up load forecasting model: forecasting individual load drivers and aggregating load post-forecast.

On screen drag & drop probabilistic AI load forecast modeling.

Build & deploy machine learning forecast models via on screen drag & drop.

Build & deploy deep learning forecast models via on screen drag & drop.

Explore our Solutions | Request a Demo | Apply for a Free Trial

AI Model Optimization

Model Splitting

Electricity data has clear daily profiles, e.g., peak, off, super peak, and similarly intra-term blocks such as weekdays and weekends; this is a total of 3 x 2 or 6 profiles. QR AI machine allows defining splitters in custom blocks or profiles. Each block is automatically assigned to its AI model, and its training data is obtained by splitting the main time series at run-time according to the selected profiles.
Model splitting amounts to the automatic creation of a series of AI models that are more robust and accurate since they are adapted to different regimes of data.

Train & Export AI Models

QR AI platform allows you to train an AI model on schedule, e.g., every night at 1.00 am, and save the trained model for execution the rest of the day, e.g., every 15 minutes, while using the latest updated predictors as new data arrives.

Features & Predictors Importance Analysis

Identifying the optimal set of predictors or features is key in optimizing AI models. During training, QR AI Forecaster performs an analysis of the predictors and ranks them by importance or effectiveness, and displays the results in a bar chart plot.
Users can iteratively add predictors, retrain the AI model and retain the best model.

Data Dashboard

Since bad input data negatively impacts an AI model, it is crucial to be able to inspect the data of each AI model, namely the target, predictors, and forecast data, before and after every stage of data processing.
All above data are gathered with plots in different sections of the dashboards for quick and easy visualization and inspection by data scientists, saving them the effort of writing code or running scripts.

QR AI Model Fine-Tuner

AI model optimization is a key step in ensuring the accuracy and the stability of forecasts. AI models can have dozens of parameters, and identifying the optimal combination that produces the best result (in terms of accuracy, shape, and behavior of forecasts) is very time-consuming and difficult. This toolbox automates the complex task of identifying the optimal set of parameters by bootstrapping a genetic computing process around a limited number of forecasts. The optimal set of parameters are those that minimize the forecast error (MAEP) over a date range. This algorithm replaces the equivalent of a hundred thousand manual trial and errors. This tool operates in two different modes: back-test and forward-test.

In back-test mode, the fine-tuner finds the best set of parameters over a past date range, e.g., one or more days.
To further improve the results of back-testing, the forward-test mode takes the optimal back-test parameters and estimates the best set of parameters that could have predicted the actual data that has become available over a range of dates.

Forecast Accuracy Analysis

We offer powerful and flexible no-coding tools to create reports and dashboards, and to visualize data.
Reports and dashboards can combine multiple data sources, e.g., actual raw data and forecast data, leading to accuracy analysis with any desired error indicator, e.g., MAPE, SMAPE, MAE.
Dashboards can be created to display side by side the errors of multiple forecast jobs with different warning thresholds for error reporting, e.g., green for good to red for high error.

Leading-Edge Electricity AI Forecasting Features

QR AI Forecaster Offers Advanced Features for Electricity Forecasting