Machine Learning Techniques For Load And Price Prediction
Expert-defined terms from the Global Certificate Course in Electricity Price Forecasting course at London School of Planning and Management. Free to read, free to share, paired with a professional course.
ARIMA (AutoRegressive Integrated Moving Average) #
ARIMA (AutoRegressive Integrated Moving Average)
Explanation #
A statistical model that captures autocorrelation in time‑series data by combining autoregressive (AR) terms, differencing (I) to achieve stationarity, and moving‑average (MA) components.
Example #
Forecasting hourly electricity demand using past demand values and residual errors.
Practical application #
Day‑ahead load prediction for utility scheduling.
Challenges #
Selecting appropriate order (p, d, q) and handling non‑linear patterns common in price series.
ARMAX (AutoRegressive Moving Average with Exogenous Inputs) #
ARMAX (AutoRegressive Moving Average with Exogenous Inputs)
Explanation #
Extends ARIMA by incorporating external regressors (e.g., temperature, fuel price) that influence the target series.
Example #
Including weather forecasts as exogenous inputs to improve load forecasts.
Practical application #
Short‑term price forecasting where market fundamentals affect prices.
Challenges #
Identifying relevant exogenous variables and avoiding multicollinearity.
Autoencoder #
Autoencoder
Explanation #
An unsupervised neural network that learns a compressed representation (encoding) of input data and reconstructs it, useful for feature extraction.
Example #
Reducing high‑dimensional smart‑meter data to a lower‑dimensional latent space before feeding into a regression model.
Practical application #
Denoising load profiles and detecting anomalies.
Challenges #
Choosing bottleneck size and preventing over‑compression that loses important patterns.
Bagging (Bootstrap Aggregating) #
Bagging (Bootstrap Aggregating)
Explanation #
Generates multiple training subsets via bootstrapping, trains a base learner on each, and aggregates predictions (usually by averaging) to improve stability.
Example #
Bagging decision trees for robust short‑term price forecasts.
Practical application #
Reducing sensitivity to outliers in load prediction.
Challenges #
Computational cost increases with number of base models.
Bayesian Optimization #
Bayesian Optimization
Explanation #
Treats hyperparameter search as a probabilistic problem, iteratively building a surrogate (often Gaussian Process) to predict performance and selecting new points via an acquisition function.
Example #
Optimizing learning rate and depth of a Gradient Boosting model for price forecasting.
Practical application #
Efficiently finding optimal model configurations with limited training runs.
Challenges #
Requires careful choice of prior and may struggle with high‑dimensional hyperparameter spaces.
Bayesian Neural Network #
Bayesian Neural Network
Explanation #
Extends conventional neural networks by placing probability distributions over weights, yielding predictive distributions rather than point estimates.
Example #
Providing confidence intervals for day‑ahead load forecasts.
Practical application #
Risk‑aware decision making in market bidding.
Challenges #
Computationally intensive training and inference.
Boosting #
Boosting
Explanation #
Sequentially trains weak learners, each focusing on errors of its predecessor, and combines them into a strong predictor.
Example #
Using XGBoost to predict hourly electricity prices based on historical price, demand, and weather features.
Practical application #
Capturing complex non‑linear relationships in price data.
Challenges #
Prone to overfitting if not regularized; requires careful early stopping.
CatBoost #
CatBoost
Explanation #
A gradient‑boosting library that handles categorical variables natively and reduces prediction shift via ordered boosting.
Example #
Incorporating fuel type categories (coal, gas, renewable) directly into a price forecasting model.
Practical application #
Faster development cycles when many categorical market features exist.
Challenges #
Memory usage can be high for very large datasets.
Clustering #
Clustering
Explanation #
Unsupervised technique that groups similar observations based on distance metrics, often used for pattern discovery.
Example #
Segmenting customers into load‑profile groups to build tailored forecasting models.
Practical application #
Reducing model complexity by applying a single model per cluster.
Challenges #
Determining optimal number of clusters and handling high‑dimensional data.
Convolutional Neural Network (CNN) #
Convolutional Neural Network (CNN)
Explanation #
Neural architecture that applies convolutional filters to capture local patterns, traditionally used for images but adaptable to time‑series via 1‑D convolutions.
Example #
Learning short‑term temporal patterns in price spikes from raw price series.
Practical application #
Automated feature extraction from high‑frequency market data.
Challenges #
Requires sufficient data to avoid over‑parameterization; filter design can be non‑intuitive for non‑spatial data.
Cross‑validation #
Cross‑validation
Explanation #
Technique for estimating model performance by partitioning data into training and validation sets multiple times.
Example #
Using a rolling‑origin evaluation to assess a load forecasting model across multiple weeks.
Practical application #
Preventing over‑optimistic error estimates in model selection.
Challenges #
Standard K‑fold breaks temporal order; special time‑aware splits are needed.
Data Augmentation #
Data Augmentation
Explanation #
Generates additional training samples by applying transformations, improving model robustness, especially when data are scarce.
Example #
Adding small Gaussian noise to historic load curves to expand the training set.
Practical application #
Enhancing model generalization for rare price spikes.
Challenges #
Augmented data may not reflect true underlying dynamics, risking bias.
Data Imputation #
Data Imputation
Explanation #
Strategies to fill gaps in datasets, essential for continuous time‑series modeling.
Example #
Using linear interpolation to replace missing hourly price values due to market outages.
Practical application #
Maintaining model continuity for real‑time forecasting.
Challenges #
Imputation can introduce artificial smoothness, affecting variance estimates.
Decision Tree #
Decision Tree
Explanation #
Tree‑structured model that splits data based on feature thresholds to minimize impurity (e.g., Gini, entropy).
Example #
Predicting price direction (up/down) using a shallow tree on recent price change and demand.
Practical application #
Interpretable baseline for load and price forecasting.
Challenges #
High variance; prone to overfitting without pruning or ensemble methods.
Dimensionality Reduction #
Dimensionality Reduction
Explanation #
Techniques that transform high‑dimensional data into a lower‑dimensional space while preserving essential information.
Example #
Applying PCA to a set of weather variables to reduce multicollinearity before model training.
Practical application #
Faster training of complex models and mitigation of the curse of dimensionality.
Challenges #
Loss of interpretability; selecting the number of components is non‑trivial.
Early Stopping #
Early Stopping
Explanation #
Halts training when performance on a validation set ceases to improve, preventing over‑learning.
Example #
Stopping LSTM training after 10 epochs with no validation loss reduction.
Practical application #
Efficient training of deep networks for price forecasting.
Challenges #
Requires a reliable validation set that reflects future data distribution.
Ensemble Kalman Filter (EnKF) #
Ensemble Kalman Filter (EnKF)
Explanation #
A recursive Bayesian filter that uses an ensemble of system states to estimate the posterior distribution, suitable for non‑linear models.
Example #
Updating load forecasts in real‑time as new smart‑meter readings arrive.
Practical application #
Improving short‑term forecasts under measurement noise.
Challenges #
Requires careful ensemble size selection; computationally intensive for large systems.
Feature Engineering #
Feature Engineering
Explanation #
Process of creating informative variables from raw data to improve model performance.
Example #
Constructing a 24‑hour lag of load and a moving‑average of temperature over the past 6 hours.
Practical application #
Enhancing predictive power of simple linear models for price forecasting.
Challenges #
Risk of data leakage if future information is inadvertently used.
Feature Scaling #
Feature Scaling
Explanation #
Rescales features to a common range, facilitating model convergence, especially for gradient‑based algorithms.
Example #
Scaling price and load values to [0,1] before feeding into a neural network.
Practical application #
Stabilizing training of deep models.
Challenges #
Scaling parameters must be stored and applied consistently during inference.
Forecast Horizon #
Forecast Horizon
Explanation #
The future time interval for which predictions are generated, ranging from minutes to years.
Example #
A 24‑hour ahead forecast for day‑ahead market bidding.
Practical application #
Aligning model choice (e.g., ARIMA vs. LSTM) with horizon length.
Challenges #
Accuracy typically degrades as horizon extends; different features become relevant.
Gaussian Process (GP) #
Gaussian Process (GP)
Explanation #
A probabilistic model that defines a distribution over functions, providing mean predictions and uncertainty estimates.
Example #
Modeling the smooth component of electricity price curves with a radial‑basis‑function kernel.
Practical application #
Quantifying forecast uncertainty for risk‑aware trading.
Challenges #
Computational complexity scales cubically with data size; requires kernel selection.
Gradient Boosting Machine (GBM) #
Gradient Boosting Machine (GBM)
Explanation #
Builds an additive model of weak learners (typically decision trees) where each new tree fits the residuals of the combined previous trees.
Example #
Using GBM to predict hourly LMPs based on historical prices, load, and renewable output.
Practical application #
Capturing non‑linear interactions in price formation.
Challenges #
Sensitive to hyperparameters; may overfit without regularization.
Grid Search #
Grid Search
Explanation #
Systematically evaluates a predefined set of hyperparameter combinations to identify the best configuration.
Example #
Testing learning rates {0.01,0.05,0.1} and max depths {3,5,7} for a Random Forest.
Practical application #
Baseline approach for model optimization.
Challenges #
Computationally expensive for large search spaces; may miss optimal regions between grid points.
Explanation #
Statistical model where observable sequences are generated by underlying hidden states following Markov dynamics.
Example #
Modeling regime shifts in electricity prices (e.g., high‑volatility vs. low‑volatility periods).
Practical application #
Detecting price regime changes for dynamic strategy adjustment.
Challenges #
Determining appropriate number of hidden states; parameter estimation can be sensitive to initialization.
Hyperparameter #
Hyperparameter
Explanation #
Settings external to the learning algorithm that control model capacity, regularization, and optimization behavior.
Example #
Number of trees in a Random Forest, dropout rate in a neural network.
Practical application #
Adjusted to balance bias‑variance trade‑off.
Challenges #
No universal rule; requires systematic search and validation.
Independent Component Analysis (ICA) #
Independent Component Analysis (ICA)
Explanation #
Decomposes multivariate signals into statistically independent components, often used to isolate underlying sources.
Example #
Separating load consumption patterns from noise in aggregated smart‑meter data.
Practical application #
Enhancing feature extraction for load forecasting.
Challenges #
Assumes non‑Gaussian sources; component ordering may be ambiguous.
Kalman Filter #
Kalman Filter
Explanation #
Recursive algorithm that provides optimal estimates of hidden states in linear Gaussian systems, combining model predictions with noisy observations.
Example #
Real‑time adjustment of load forecasts as new SCADA measurements arrive.
Practical application #
Reducing forecast latency in operational environments.
Challenges #
Assumes linearity and Gaussian noise; extensions needed for non‑linear dynamics.
K‑means Clustering #
K‑means Clustering
Explanation #
Partitions data into K clusters by minimizing within‑cluster sum of squares, assigning each point to the nearest centroid.
Example #
Grouping daily load curves into typical “weekday”, “weekend”, and “holiday” shapes.
Practical application #
Building specialized models per cluster to improve accuracy.
Challenges #
Requires pre‑specifying K; sensitive to outliers and initial centroid placement.
L1 Regularization (Lasso) #
L1 Regularization (Lasso)
Explanation #
Adds the absolute value of coefficients to the loss function, encouraging many coefficients to become exactly zero.
Example #
Selecting the most influential weather variables for a linear price model.
Practical application #
Simplifying models for interpretability and reducing overfitting.
Challenges #
May arbitrarily drop correlated predictors; selection stability can vary with data splits.
L2 Regularization (Ridge) #
L2 Regularization (Ridge)
Explanation #
Adds the squared magnitude of coefficients to the loss function, shrinking them toward zero but rarely eliminating them entirely.
Example #
Stabilizing a multivariate regression of price on demand and fuel costs.
Practical application #
Improves generalization when many correlated features exist.
Challenges #
Does not produce sparse solutions; selecting regularization strength is critical.
LightGBM #
LightGBM
Explanation #
An efficient gradient‑boosting framework that grows trees leaf‑wise and uses histogram binning for speed and memory savings.
Example #
Forecasting 15‑minute ahead prices with millions of observations.
Practical application #
Real‑time price prediction where latency matters.
Challenges #
Can overfit on small datasets; requires careful tuning of leaf‑wise parameters.
Long Short‑Term Memory (LSTM) #
Long Short‑Term Memory (LSTM)
Explanation #
A type of RNN that introduces input, forget, and output gates to preserve long‑range dependencies while mitigating vanishing gradients.
Example #
Modeling daily load patterns with an LSTM that ingests past 168 hours of load and weather.
Practical application #
Accurate multi‑step ahead forecasts for day‑ahead markets.
Challenges #
Requires substantial training data; hyperparameter selection (layers, units) impacts performance.
MAE (Mean Absolute Error) #
MAE (Mean Absolute Error)
Explanation #
Average of absolute differences between predicted and actual values, providing an intuitive measure of average error magnitude.
Example #
MAE of 2.3 MW for a 24‑hour load forecast.
Practical application #
Benchmarking models where penalizing large errors is less critical than overall bias.
Challenges #
Does not penalize large outliers as heavily as squared‑error metrics.
MAPE (Mean Absolute Percentage Error) #
MAPE (Mean Absolute Percentage Error)
Explanation #
Mean of absolute percentage differences, expressing error as a proportion of actual values.
Example #
MAPE of 4 % for hourly price forecasts.
Practical application #
Comparing models across different scales (e.g., low‑load vs. high‑load periods).
Challenges #
Undefined when actual values approach zero; can be skewed by small denominators.
Monte Carlo Simulation #
Monte Carlo Simulation
Explanation #
Generates a large number of random scenarios based on probability distributions to assess uncertainty and risk.
Example #
Simulating 10 000 possible price paths for a given day using estimated volatility.
Practical application #
Estimating Value‑at‑Risk (VaR) for market participants.
Challenges #
Computationally heavy; results depend on assumed distributions.
Multi‑Layer Perceptron (MLP) #
Multi‑Layer Perceptron (MLP)
Explanation #
A classic neural network consisting of an input layer, one or more hidden layers, and an output layer, where each neuron applies a weighted sum followed by a non‑linear activation.
Example #
Predicting next‑hour price using an MLP with ReLU activation and three hidden layers.
Practical application #
Baseline deep learning model for quick prototyping.
Challenges #
Limited ability to capture temporal dependencies without engineered lag features.
Multicollinearity #
Multicollinearity
Explanation #
Situation where two or more explanatory variables are highly linearly related, inflating variance of coefficient estimates.
Example #
Temperature and humidity often exhibit multicollinearity in weather‑driven load models.
Practical application #
Diagnosing and remedying via feature selection or PCA.
Challenges #
Can obscure true predictor importance and destabilize regression models.
Neural Architecture Search (NAS) #
Neural Architecture Search (NAS)
Explanation #
Automated process that explores the space of neural network architectures to identify high‑performing designs.
Example #
Using NAS to discover an optimal combination of convolutional and recurrent layers for price forecasting.
Practical application #
Reducing manual effort in designing deep models.
Challenges #
Extremely computationally expensive; risk of overfitting to validation set.
Normalization #
Normalization
Explanation #
Adjusts data to have a particular statistical property (commonly zero mean and unit variance), facilitating model training.
Example #
Normalizing load values before feeding them into a Support Vector Machine.
Practical application #
Improves convergence speed of gradient‑based optimizers.
Challenges #
Must apply same transformation to future data; outliers can affect mean and variance.
Online Learning #
Online Learning
Explanation #
Model updates occur continuously as new data arrive, without retraining from scratch.
Example #
Updating a price prediction model hourly as market data streams in.
Practical application #
Maintaining up‑to‑date forecasts in fast‑changing markets.
Challenges #
Balancing stability and adaptability; handling noisy updates.
Outlier Detection #
Outlier Detection
Explanation #
Identifies observations that deviate markedly from the majority, which may indicate data errors or rare events.
Example #
Flagging a sudden 30 % price jump as a potential market anomaly.
Practical application #
Cleaning training data to avoid biasing models.
Challenges #
Distinguishing genuine extreme events from erroneous measurements.
Overfitting #
Overfitting
Explanation #
When a model captures noise in training data, leading to poor generalization on unseen data.
Example #
A deep LSTM with many layers that perfectly fits historical price but fails on new days.
Practical application #
Recognizing and mitigating via validation, early stopping, or dropout.
Challenges #
Detecting overfitting early in time‑series contexts where validation data may be limited.
Parametric Model #
Parametric Model
Explanation #
Model defined by a finite set of parameters with a predetermined functional form.
Example #
A linear regression of price on load and fuel cost with fixed coefficients.
Practical application #
Simpler interpretation and faster training.
Challenges #
May be too rigid to capture complex non‑linear relationships.
Partial Autocorrelation Function (PACF) #
Partial Autocorrelation Function (PACF)
Explanation #
Measures correlation between a time series and its lag after removing the effects of intermediate lags, aiding in AR order identification.
Example #
PACF spikes at lag 24 suggest a daily seasonality component in load data.
Practical application #
Guiding ARIMA model specification.
Challenges #
Requires stationary series; noisy data can produce ambiguous spikes.
Peak Load #
Peak Load
Explanation #
The maximum electricity demand observed over a specific period, often dictating system capacity requirements.
Example #
A 45 GW peak observed during a summer afternoon.
Practical application #
Forecasting peaks to schedule generation and manage market pricing.
Challenges #
Peaks are rare and highly sensitive to weather and behavioral factors.
Permutation Importance #
Permutation Importance
Explanation #
Assesses the impact of each feature by randomly permuting its values and measuring the resulting drop in model performance.
Example #
Determining that temperature has the highest permutation importance for load forecasts.
Practical application #
Guiding feature selection and model explanation.
Challenges #
Computationally expensive for large datasets; can be biased for correlated features.
Pinball Loss #
Pinball Loss
Explanation #
Measures accuracy of quantile forecasts by penalizing under‑ and over‑predictions asymmetrically according to the quantile level.
Example #
Pinball loss of 0.12 for the 0.9‑quantile price forecast.
Practical application #
Evaluating probabilistic price forecasts required for risk management.
Challenges #
Requires multiple quantile predictions; interpretation can be less intuitive than MSE.
Power Law Distribution #
Power Law Distribution
Explanation #
Statistical distribution where the probability of an event scales as a power of its size, often observed in price spikes.
Example #
Modeling the tail of price spikes with a Pareto distribution.
Practical application #
Risk assessment for extreme price events.
Challenges #
Parameter estimation is sensitive to sample size; tail behavior may change over time.
Probabilistic Forecasting #
Probabilistic Forecasting
Explanation #
Generates a full distribution (or set of quantiles) for future values rather than a single point estimate.
Example #
Producing the 5th, 50th, and 95th percentile forecasts for hourly load.
Practical application #
Enabling market participants to hedge against uncertainty.
Challenges #
Requires more sophisticated evaluation metrics (e.g., CRPS) and larger training data.
Principal Component Analysis (PCA) #
Principal Component Analysis (PCA)
Explanation #
Linear technique that transforms correlated variables into a set of uncorrelated components ordered by explained variance.
Example #
Reducing 10 weather variables to 3 principal components before regression.
Practical application #
Mitigating multicollinearity and speeding up training.
Challenges #
Components may lack physical interpretability; linear assumption may miss non‑linear structures.
Random Forest #
Random Forest
Explanation #
Ensemble of decorrelated decision trees built on bootstrapped samples, with random feature selection at each split, producing robust predictions.
Example #
Predicting day‑ahead price using a Random Forest with 500 trees.
Practical application #
Handles mixed data types and provides built‑in importance scores.
Challenges #
Can be memory intensive; less effective for extrapolation beyond training range.
Recurrent Neural Network (RNN) #
Recurrent Neural Network (RNN)
Explanation #
Neural architecture designed for sequential data, where each step’s output depends on previous hidden state.
Example #
Modeling hourly price sequences with a simple RNN layer.
Practical application #
Capturing short‑term temporal patterns.
Challenges #
Training instability due to vanishing/exploding gradients; LSTM/GRU often preferred.
Regularization #
Regularization
Explanation #
Techniques that add constraints or penalties to the loss function to prevent overfitting and improve generalization.
Example #
Applying L2 regularization to the weights of a neural network for price forecasting.
Practical application #
Stabilizing models with limited training data.
Challenges #
Selecting appropriate regularization strength; overly strong regularization may underfit.
Residual Analysis #
Residual Analysis
Explanation #
Examination of the differences between observed and predicted values to assess model adequacy.
Example #
Detecting remaining seasonal patterns in residuals of an ARIMA load model.
Practical application #
Guiding model refinement and identifying missing dynamics.
Challenges #
Requires sufficient residual data; statistical tests may have limited power for short series.
Ridge Regression #
Ridge Regression
Explanation #
Linear regression variant that penalizes the sum of squared coefficients, reducing variance at the cost of some bias.
Example #
Stabilizing price prediction coefficients when fuel price and carbon price are highly correlated.
Practical application #
Improves out‑of‑sample performance for high‑dimensional feature sets.
Challenges #
Does not perform variable selection; all predictors remain in the model.
Rolling Window Validation #
Rolling Window Validation
Explanation #
Sequentially moves a fixed‑size training window forward in time, training and testing at each step to mimic real‑world forecasting.
Example #
Using a 30‑day training window to predict the next 24 hours, then rolling forward by one day.
Practical application #
Provides realistic performance estimates for operational models.
Challenges #
Computationally demanding; choice of window size influences bias‑variance trade‑off.
Scenario Analysis #
Scenario Analysis
Explanation #
Evaluates model performance under a set of predefined alternative future conditions.
Example #
Assessing load forecasts under a heatwave scenario versus a normal summer day.
Practical application #
Planning generation capacity and market strategies under extreme conditions.
Challenges #
Defining realistic and comprehensive scenarios; may require additional data sources.
Seasonality #
Seasonality
Explanation #
Regular, repeating fluctuations in a time series occurring at fixed intervals (e.g., daily, weekly, yearly).
Example #
Higher electricity demand every weekday morning due to work‑day routines.
Practical application #
Incorporating seasonal components in ARIMA or neural network models to improve accuracy.
Challenges #
Seasonality can shift over time (e.g., due to policy changes), requiring adaptive modeling.
Shapley Additive Explanations (SHAP) #
Shapley Additive Explanations (SHAP)
Explanation #
Provides consistent, locally accurate attribution values for each feature based on cooperative game theory.
Example #
Explaining that a sudden temperature rise contributed +0.8 MW to the load forecast for a specific hour.
Practical application #
Building trust with stakeholders by elucidating black‑box model decisions.
Challenges #
Computationally intensive for large models; approximations may be needed.
Signal‑to‑Noise Ratio (SNR) #
Signal‑to‑Noise Ratio (SNR)
Explanation #
Ratio of the power of the desired signal to the power of background noise, indicating data reliability.
Example #
High SNR in SCADA measurements versus lower SNR in low‑voltage smart‑meter data.
Practical application #
Determining appropriate preprocessing steps (e.g., smoothing) before model training.
Challenges #
Estimating noise characteristics in real‑world electricity data can be difficult.
Simple Exponential Smoothing (SES) #
Simple Exponential Smoothing (SES)
Explanation #
Forecasting method that applies exponentially decreasing weights to past observations, suitable for data without trend or seasonality.
Example #
Using SES to predict short‑term price when the market is relatively stable.
Practical application #
Baseline comparison for more complex models.
Challenges #
Cannot capture trends or seasonal patterns; smoothing parameter selection is critical.
Single‑Step Forecast #
Single‑Step Forecast
Explanation #
Predicts the immediate next value in a series, often used as a building block for multi‑step forecasts.
Example #
Forecasting the price for the next 15‑minute interval.
Practical application #
Real‑time bidding decisions in intraday markets.
Challenges #
Errors can accumulate when iteratively feeding predictions back into the model for longer horizons.
Spatial‑Temporal Modeling #
Spatial‑Temporal Modeling
Explanation #
Incorporates both spatial relationships (e.g., between nodes in a grid) and temporal dynamics into a unified framework.
Example #
Predicting nodal LMPs by considering neighboring bus prices and historical trends.
Practical application #
Enhancing location‑specific price forecasts for congestion management.
Challenges #
High dimensionality; requires extensive data on network topology.
State‑Space Model #
State‑Space Model
Explanation #
Represents a system with an unobserved state vector evolving over time and observed measurements linked to that state.
Example #
Modeling the hidden demand trend and observed load simultaneously.
Practical application #
Provides a flexible framework for incorporating both observed and latent variables.
Challenges #
Model identification can be complex; requires assumptions about noise distributions.
Stochastic Gradient Descent (SGD) #
Stochastic Gradient Descent (SGD)
Explanation #
Iterative optimization algorithm that updates model parameters using gradients computed on small random subsets of data.
Example #
Training a large‑scale price prediction neural network with SGD and momentum.
Practical application #
Enables training on massive datasets that cannot fit into memory.
Challenges #
Sensitive to learning‑rate schedule; may converge to noisy minima.
Support Vector Machine (SVM) #
Support Vector Machine (SVM)
Explanation #
Supervised learning algorithm that finds the hyperplane that maximally separates classes (classification) or fits within a tolerance (regression).
Example #
Using ε‑SVR to predict price deviations from a baseline.
Practical application #
Effective for small‑to‑medium datasets with non‑linear relationships when combined with kernels.
Challenges #
Scaling to large datasets is difficult; kernel choice heavily influences performance.
Temporal Fusion Transformers (TFT) #
Temporal Fusion Transformers (TFT)
Explanation #
Advanced deep learning architecture that combines recurrent layers, attention, and gating mechanisms to handle both static and time‑varying inputs for multi‑step forecasts.
Example #
Forecasting 24‑hour ahead prices using recent market data, calendar features, and weather forecasts.
Practical application #
State‑of‑the‑art performance on complex electricity price series.
Challenges #
Requires substantial computational resources; interpretability is less straightforward than classical models.
Time‑Series Cross‑validation #
Time‑Series Cross‑validation
Explanation #
Validation technique that respects temporal order by training on past data and testing on future data repeatedly.
Example #
Using a 12‑month training window to predict the following month, then shifting forward month by month.
Practical application #
Provides realistic performance estimates for forecasting models.
Challenges #
Limited number of folds for short series; may still leak information if seasonal patterns span folds.
Transfer Learning #
Transfer Learning
Explanation #
Leverages knowledge learned from a source task to improve performance on a related target task, often by reusing model weights.
Example #
Pre‑training a neural network on large‑scale European price data, then fine‑tuning on a specific regional market.
Practical application #
Reduces data requirements for new markets with limited historical records.
Challenges #
Mismatch between source and target domains can lead to negative transfer.
Trend Component #
Trend Component
Explanation #
Underlying direction (increasing or decreasing) in a time series, separate from seasonal and irregular components.
Example #
A gradual rise in average electricity prices due to policy changes.
Practical application #
Removing trend before fitting stationary models like ARIMA.
Challenges #
Trend may be non‑linear; detecting it reliably requires sufficient data length.
t‑Distributed Stochastic Neighbor Embedding (t‑SNE) #
t‑Distributed Stochastic Neighbor Embedding (t‑SNE)
Explanation #
Non‑linear dimensionality reduction technique that maps high‑dimensional data to a low‑dimensional space while preserving local structure.
Example #
Visualizing clusters of load profiles to identify typical consumption patterns.
Practical application #
Exploratory data analysis for feature engineering.
Challenges #
Computationally intensive; results can vary with hyperparameters, making reproducibility an issue.
Underfitting #
Underfitting
Explanation #
When a model is too simple to capture underlying patterns, leading to poor performance on both training and validation data.
Example #
Using a linear regression to model highly non‑linear price spikes.
Practical application #
Diagnosing model inadequacy and prompting the use of more expressive algorithms.
Challenges #
Balancing model complexity to avoid swinging into overfitting.
Vanilla RNN #
Vanilla RNN