Multivariate LSTM-Based Forecasting for Renewable Energy: Enhancing Climate Change Mitigation

AI summary

Overview

• Research area: Machine learning for power systems — specifically multivariate deep learning (LSTM) applied to renewable energy generation forecasting, linked to power system economic dispatch and climate change mitigation.
• Technical level: Intermediate. The paper assumes some familiarity with neural networks (LSTM units, activation functions, dropout, loss functions) and with power system concepts such as day-ahead and real-time economic dispatch, load shedding, and renewable spillage.
• Scope: The paper proposes a multivariate LSTM network that forecasts solar photovoltaic (PV) generation across three planning areas in Alberta using one year of hourly historical data, and evaluates the forecast's downstream impact on CO2 emissions, load shedding, spillage, and dispatch cost.

What This Paper Is About

Renewable energy sources such as solar and wind vary over time, and grid operators need accurate forecasts of how much power they will produce in order to balance supply and demand reliably and cheaply. Traditional approaches, including deterministic methods and stochastic programming that relies on K-means clustering to build representative days, often fail to capture the complex temporal dependencies and non-linear patterns in renewable generation data. This paper builds a multivariate LSTM model that forecasts PV generation for a target area by using historical data from that area plus neighboring areas, and then tests whether better forecasts translate into lower emissions, less load shedding, and lower operating cost.

Key Contributions

1. A multivariate LSTM (M-LSTM) forecasting network that captures latent features and long-term dependencies among renewable energy sources across different regions, predicting generation for one area using its own historical data plus data from neighboring areas.
2. A formulation that accounts for the temporal gap between day-ahead market closure and the start of the real-time market when forecasting renewable generation.
3. An economic dispatch case study coupling the forecasts to a day-ahead and real-time dispatch problem with three synchronous generators and one PV source, quantifying how forecast quality affects CO2 emissions, load shedding, renewable spillage, and total cost.
4. A benchmark comparison of the M-LSTM against a K-means clustering approach and a monthly-average-per-hour approach, using Normalized Mean Absolute Error (NMAE) under Mean Normalization as the accuracy metric.

Main Findings

• Lower emissions with M-LSTM: CO2 emissions were 53,267 kg in Case 1 (K-means), 28,393 kg in Case 2 (monthly average), and 21,957 kg with the M-LSTM — which the authors describe as a 58% decrease relative to Case 1.
• Reduced reliance on gas-fired generation: Gas-fired unit output was 263.7 MW in Case 1, 140.56 MW in Case 2, and 108.7 MW with M-LSTM.
• Elimination of load shedding: Load shedding was 121.7 MW in Case 1 and 5.79 MW in Case 2, and dropped to 0 MW with M-LSTM.
• Slightly higher renewable spillage: RES spillage was 0 MW in Case 1, 0 MW in Case 2, and 5 MW with M-LSTM. The authors attribute this to the M-LSTM's superior forecast accuracy — other models avoid spillage by overestimating renewable generation, which masks their forecast errors.
• Lowest dispatch cost: DA+RT cost was $53,043 in Case 1, $44,148 in Case 2, and $43,490 with M-LSTM.
• Best forecast accuracy: NMAE was 1.57 (Case 1), 0.59 (Case 2), and 0.46 (M-LSTM).
• Balanced trade-off: The authors argue the minimal spillage under M-LSTM reflects optimal system balancing, maximizing renewable usage while maintaining grid stability.

Methodology in Plain English

The researchers took one full year of real hourly renewable generation data from Alberta, covering a single resource type (PV) across three different planning areas. The dataset contains 8,760 data samples per feature, with three features in total.

Network design. The forecasting model is a two-layer LSTM network. The first layer has 64 LSTM units with a ReLU activation function; the second has 32 LSTM units, also with ReLU. Outputs from the second LSTM layer pass through a dropout layer to prevent overfitting, then into a fully connected neural network that produces the final predictions.

Training setup. The inputs are sequences of data points from time step 1 to time step p (the number of look-back steps) across all features, and a shifted sequence from step 2 to step p+1. The corresponding labels are values at p+m and p+m+1 for the feature being predicted, where m is the number of steps ahead. Input data was normalized to prevent large-scale features from dominating learning, and the network used the Adam optimizer with Mean Squared Error as the loss function. Outputs were rescaled back to their original scale after training.

Domain knowledge injection. The authors hard-coded prior knowledge about dark hours: during predefined nighttime periods, PV generation was explicitly set to zero, removing the need for predictions during those hours.

Test setup. The model looks back at the past 24 hours of PV generation across all three regions to predict generation for the target time step in the desired source. The forecasts feed a day-ahead economic dispatch, and then a real-time economic dispatch is run using the actual renewable values that become available. The gap between forecast and actual is covered by conventional generators, assumed to emit 202 kg CO2 per MWh. The test system uses three synchronous generators (G1: 20 $/MW cost, 50 MW max, 20 MW/h ramp; G2: 25 $/MW, 50 MW max, 20 MW/h ramp; G3: 30 $/MW, 30 MW max, 30 MW/h ramp, available in both DA and RT) plus one PV source.

Why This Matters

Impact on research. The paper connects forecasting accuracy directly to measurable grid outcomes — emissions, load shedding, spillage, and cost — rather than reporting forecast error alone. It also argues that the choice of benchmark matters: a forecast that appears to avoid spillage entirely may simply be overestimating renewable output, hiding its own error.

Real-world applications.

• Day-ahead and real-time market scheduling by grid operators and system planners.
• Renewable portfolio and dispatch optimization for utilities managing solar fleets across multiple regions.
• Carbon accounting and policy design, since the paper's framing connects dispatch behavior to marginal emissions from flexible gas generation.
• Regional coordination: using data from neighboring areas to improve local forecasts, which is relevant for interconnected power systems.

Industry relevance. The 58% emissions reduction reported from Case 1 to the M-LSTM case, alongside the elimination of load shedding and a lower total dispatch cost, suggests that AI-driven forecasting could support both decarbonization targets and grid reliability. The authors note that scaled deployment across power generation systems could significantly reduce CO2 emissions.

Future Directions

• Extension to other renewable types: The study used only PV across three planning areas; the conclusion describes the model as capable of capturing dependencies across multiple locations and RES types, but wind and other resources were not tested here.
• Scaled deployment: The authors state that results demonstrate the potential of scaled use of the M-LSTM across power generation systems, leaving the question of how the approach performs at larger geographic and system scale open.
• The spillage trade-off: The 5 MW of spillage under M-LSTM versus 0 MW under the other models raises the question of how to value accurate-but-slightly-over-supplied forecasts against conservative forecasts that hide error, especially under different market rules.
• Temporal-gap modeling: The paper highlights the gap between day-ahead market closure and real-time market start as a consideration; how best to exploit that gap with additional data or updated forecasts is not resolved.

Target Audience

This paper is most useful for power system operators, energy market analysts, and researchers working on renewable integration and grid decarbonization who want to see forecasting accuracy tied to concrete operational and emissions outcomes. It also suits machine learning practitioners interested in applying recurrent architectures to multivariate time series with spatial correlation, and policy analysts interested in how marginal generation from flexible gas units shapes the emissions impact of renewables. Readers without background in LSTM networks or economic dispatch formulations will find the appendices and modeling details dense, though the headline comparison table is accessible.

Not reported in this paper: training hyperparameters beyond the optimizer and loss function (for example, learning rate, batch size, dropout rate, number of epochs), the exact calendar period of the data, train/validation/test splits, the specific value of the look-back parameter p other than the stated 24-hour look-back, and the number of steps ahead m.

Authors’ abstract

The increasing integration of renewable energy sources (RESs) into modern power systems presents significant opportunities but also notable challenges, primarily due to the inherent variability of RES generation. Accurate forecasting of RES generation is crucial for maintaining the reliability, stability, and economic efficiency of power system operations. Traditional approaches, such as deterministic methods and stochastic programming, frequently depend on representative scenarios generated through clustering techniques like K-means. However, these methods may fail to fully capture the complex temporal dependencies and non-linear patterns within RES data. This paper introduces a multivariate Long Short-Term Memory (LSTM)-based network designed to forecast RESs generation using their real-world historical data. The proposed model effectively captures long-term dependencies and interactions between different RESs, utilizing historical data from both local and neighboring areas to enhance predictive accuracy. In the case study, we showed that the proposed forecasting approach results in lower CO2 emissions, and a more reliable supply of electric loads.