LSTM-GA-SHAP Hybrid Forecasting Approach

• The LSTM-GA-SHAP approach is a hybrid workflow combining LSTM for temporal modeling, GA for optimization, and SHAP for feature attribution to enhance prediction accuracy.
• It enables modular integration where each component addresses specific tasks—capturing sequence dependencies, tuning hyperparameters, and guiding feature selection across various applications.
• Empirical studies highlight significant improvements in error metrics and model interpretability, while also noting challenges like recurrent explainability and computational overhead.

Searching arXiv for the cited papers to ground the article. arXiv search query: (Bhupatiraju et al., 29 Jul 2025) Explainability-Driven Feature Engineering for Mid-Term Electricity Load Forecasting in ERCOT's SCENT Region The LSTM-GA-SHAP approach denotes a family of forecasting, optimization, and explanation pipelines that combine Long Short-Term Memory (LSTM) models, genetic algorithm (GA) procedures, and Shapley Additive Explanations (SHAP). In the recent literature, however, the label does not identify a single standardized architecture. Some papers implement all three components directly, using LSTM for sequence modeling, GA for hyperparameter tuning or operational optimization, and SHAP for feature selection or post hoc attribution; other papers use the label more loosely, with no explicit GA at all, or with SHAP applied only to surrogate or non-recurrent models rather than to the final LSTM itself (Rahimi et al., 8 Jul 2025, Alba et al., 25 May 2026, Bhupatiraju et al., 29 Jul 2025).

1. Terminological scope and literature usage

A concise way to understand the term is to treat it as an umbrella for hybrid sequential-learning workflows in which temporal prediction, search-based optimization, and additive attribution are combined in different proportions. The literature shows that the name is heterogeneous rather than canonical.

Paper	Domain	Configuration actually used
"Explainability-Driven Feature Engineering for Mid-Term Electricity Load Forecasting in ERCOT's SCENT Region" (Bhupatiraju et al., 29 Jul 2025)	Mid-term electricity load forecasting	SHAP-guided feature engineering; no explicit GA
"AI-Based Demand Forecasting and Load Balancing for Optimising Energy use in Healthcare Systems: A real case study" (Rahimi et al., 8 Jul 2025)	Hospital energy forecasting and load balancing	LSTM forecasting, GA tuning/load allocation, SHAP via surrogate models
"Electricity Consumption Forecasting: An Approach Using Cooperative Ensemble Learning with SHapley Additive exPlanations" (Alba et al., 25 May 2026)	University-campus electricity consumption	SHAP feature selection, GA optimization, GA-LSTM within WSB

The ERCOT load-forecasting study is especially important for clarifying nomenclature. It explicitly states that there is no explicit genetic algorithm (GA) in the methodology, no GA-based feature-selection loop, and no GA-based hyperparameter optimization; the method is instead an iterative, SHAP-guided refinement cycle with feature ablation and retraining (Bhupatiraju et al., 29 Jul 2025). By contrast, the healthcare energy paper presents an integrated LSTM-GA-SHAP framework in which GA tunes model parameters and optimizes load allocation, while SHAP explains lagged-input contributions (Rahimi et al., 8 Jul 2025). The campus electricity paper likewise contains a literal GA component, using GA for hyperparameter optimization and SHAP for feature selection before cooperative ensembling (Alba et al., 25 May 2026).

Related studies widen the term further. The recession-prediction paper uses LSTM and GRU with SHAP, but not GA (Chung, 2023). The American option-pricing paper is even farther from the literal label: its flagship model is Self-Attention GRU, interpreted with SHAP, and it explicitly notes that describing it as “LSTM-GA-SHAP” would be misleading (Shen, 2023). This suggests that the expression often functions as a shorthand for a broader class of sequence-model-plus-explainability systems rather than a fixed algorithmic template.

2. Core components and formal structure

At the modeling core, the LSTM component is used for sequential dependence, lag structure, and stateful temporal representation. In the ERCOT forecasting study, the standard LSTM cell is written as

$i_t = \sigma(W_i x_t + U_i h_{t-1} + b_i)$

$\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$

$f_t = \sigma(W_f x_t + U_f h_{t-1} + b_f)$

$C_t = f_t \odot C_{t-1} + i_t \odot \tilde{C}_t$

$o_t = \sigma(W_o x_t + U_o h_{t-1} + b_o)$

$h_t = o_t \odot \tanh(C_t),$

with the usual interpretation of input, forget, and output gates, candidate state, cell state, and hidden state (Bhupatiraju et al., 29 Jul 2025).

The SHAP component is grounded in the Shapley-value formulation. The ERCOT paper defines the attribution for feature $i$ as

$$\phi_i = \sum_{S \subseteq F \setminus \{i\}} \frac{|S|! (|F| - |S| - 1)!}{|F|!} \left[ f(S \cup \{i\}) - f(S) \right],$$

so that explanation is the average marginal contribution of a feature across coalitions of the remaining features (Bhupatiraju et al., 29 Jul 2025). The campus electricity paper uses the same conceptual basis and the additive decomposition

$f(x) = \phi_0 + \sum_i \phi_i,$

with $\phi_0$ as the baseline and $\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$0 as the feature contribution (Alba et al., 25 May 2026).

The GA component is not standardized across studies. In the healthcare energy framework, the LSTM-tuning chromosome is

$\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$1

and the fitness is defined as negative MAE,

$\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$2

so the GA maximizes fitness by minimizing prediction error (Rahimi et al., 8 Jul 2025). In the same paper, load balancing is also cast as minimizing mismatch between allocation and forecast,

$\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$3

In the campus electricity study, GA optimizes hyperparameters for LSTM, RF, SVR, and XGBoost, with the explicit objective of minimizing MAE under 5-fold cross-validation (Alba et al., 25 May 2026).

A methodological complication is that recurrent models are not trivially explainable with feed-forward attribution machinery. The LRP study on LSTM interpretation argues that LSTMs contain linear mappings, gated multiplicative interactions, and accumulation, and therefore cannot be explained faithfully by naively applying standard feed-forward rules (Arras et al., 2019). This technical point is directly relevant to why some “LSTM-GA-SHAP” workflows deploy SHAP on surrogate or tree models rather than on the recurrent network itself.

3. Recurrent workflow patterns

One major pattern is explainability-driven feature engineering. The ERCOT study uses hourly load data from ERCOT’s SCENT region and hourly NOAA weather data, aggregated from Austin, San Antonio, Round Rock, and San Marcos using a population-weighted average. Data are aligned to Central Time, resampled hourly, corrected for daylight saving time and leap years, cleaned by linear interpolation and forward fill, and split chronologically into 2016–2023 for training and 2024 for testing. The methodological loop is explicit: train a baseline model, compute SHAP values, inspect where predictions fail—especially during peak-load windows—engineer new variables, retrain, recompute SHAP, and reassess both global and peak-specific accuracy (Bhupatiraju et al., 29 Jul 2025).

The engineered variables in that pipeline include calendar, lagged-load, rolling, spike, interaction, and thermal-extreme features such as load_spike_vs_mean, load_roll_mean_24, load_roll_max_24, load_roll_std_24, load_roll_mean_336, load_roll_std_168, load_lag_48, load_lag_168, tmax_lag_24, tmin_lag_24, tmax_roll_max_72, [CDD](https://www.emergentmind.com/topics/continuous-dynamical-decoupling-cdd), CDD_x_hour, lag_24_x_hour, temp_spike_vs_mean, is_extreme_heat_event, is_extreme_cold_event, and cyclical encodings like hour_sin and dayofweek_cos. Representative formulas include

$\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$4

$\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$5

and

$\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$6

The LSTM in this study uses the refined inputs indirectly, because SHAP was not used for LSTM: the paper states that Deep SHAP / gradient-based explainers were unstable for the recurrent architecture (Bhupatiraju et al., 29 Jul 2025).

A second pattern is forecasting plus optimization. In the healthcare case study, hourly hospital energy data are scaled with MinMaxScaler to $\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$7, converted into supervised sequences with a sliding window, and structured so that the past 24 hours predict the next hour. The recurrent network contains two LSTM layers, dropout after each LSTM layer, a dense layer with 25 neurons, and a final dense output with 1 neuron. GA is then used both for LSTM hyperparameter tuning and for load balancing, with tournament selection, SBX crossover, adaptive mutation with Gaussian noise, and replacement of the worst individuals by offspring (Rahimi et al., 8 Jul 2025).

A third pattern is feature selection plus cooperative ensembling. The campus electricity paper begins with a candidate set that includes 12 lags of electricity consumption, climate variables, and calendar/pandemic indicators over 79 months from September 2017 to March 2024. SHAP importances are computed from RF and XGBoost on the Palmas dataset, averaged, and used in sequential elimination. Removing 36 least-important features minimizes average MAE and leaves 11 final predictors: Lag 1, Lag 2, Lag 3, Lag 6, Lag 12, Maximum temperature – average absolute, Minimum temperature – absolute, Maximum wind speed – absolute, Month = March, Year = 2019, and COVID-19 identifier. GA then tunes the main machine-learning models, and the final Weaker Separator Booster (WSB) uses GA-LSTM as the strong predictor and the mean of GA-RF, GA-SVR, and GA-XGBoost as the weaker reference (Alba et al., 25 May 2026).

4. Empirical behavior across application domains

In electricity load forecasting for ERCOT, the reported results show large differences between baseline and SHAP-informed variants. The performance table gives MAPE = 0.79\% for XGBoost (Improved with SHAP), 0.91\% for LightGBM (Improved with SHAP), 1.21\% for LSTM, and 1.14\% for Bidirectional LSTM. Linear regression improves with engineered features but remains the weakest family. The paper also states that peak MAPE reportedly fell by about 3–6x, with the largest gains during peak periods, including winter extremes, summer extremes, and the top 5\% of load hours (Bhupatiraju et al., 29 Jul 2025).

In the healthcare energy case study, the comparison is against ARIMA and Prophet. The reported forecasting results are ARIMA: MAE 87.73, RMSE 125.22, Prophet: MAE 59.78, RMSE 81.22, and LSTM: MAE 21.69, RMSE 29.96. The paper interprets this as evidence that ARIMA struggles with non-linear, highly dynamic demand, Prophet improves through seasonality and trend handling, and LSTM best captures complex temporal patterns (Rahimi et al., 8 Jul 2025).

In the university-campus electricity study, the key 12-month-ahead test results are campus-specific. For IFPR-Palmas, WSB achieves sMAPE 13.90\% and MAE 1990.87 kWh, outperforming GA-LSTM at sMAPE 15.35\% and MAE 2264.50 kWh. For IFPR-Coronel Vivida, WSB achieves sMAPE 18.72\% and MAE 464.93 kWh, slightly better than GA-LSTM at sMAPE 18.80\% and MAE 466.92 kWh. The same paper states that LSTM was the best standalone model in the 12-month-ahead test phase, while WSB improved further (Alba et al., 25 May 2026).

Outside energy systems, the same component pattern appears in other domains. In real-time US recession prediction, LSTM and GRU outperform standard linear models, particularly at longer horizons and under class imbalance, with SHAP identifying predictors such as the S&P 500 index at short horizons and the term spread at longer horizons (Chung, 2023). In American option pricing, the best model is Self-Attention GRU with 21 features, and SHAP reveals that spot price1, strike1, and moneyness1 are the most influential features, with historical call prices also strongly positive (Shen, 2023). These studies show that the LSTM-GA-SHAP label, or adjacent variants of it, has expanded well beyond one application area.

5. Explainability, attribution, and recurrent-specific caveats

SHAP serves different roles across the literature. It is used as an interpretability tool, a feature-engineering guide, a feature-selection mechanism, and, in some cases, a surrogate explanation layer for recurrent models that are not directly explained. In the ERCOT study, SHAP is run after model training to identify which variables drive predictions, which are weak or redundant, and where the model underpredicts peak demand. The paper emphasizes that baseline models often over-relied on month, dayofweek, and hour, while the stronger predictors for difficult demand events were recent load history, lagged/weather interactions, and spike/volatility indicators (Bhupatiraju et al., 29 Jul 2025).

The recurrent setting introduces a more difficult attribution problem than static tabular modeling. The LRP study argues that an LSTM contains gated products such as signal $\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$8 gate and accumulators of the form

$\tilde{C}_t = \tanh(W_C x_t + U_C h_{t-1} + b_C)$9

so explanation must propagate through multiplicative control and memory recursion rather than only across feed-forward layers. Its preferred rule for gated interactions is LRP-all,

$f_t = \sigma(W_f x_t + U_f h_{t-1} + b_f)$0

motivated by Deep Taylor Decomposition and by the semantic idea that the signal, rather than the gate, should receive relevance (Arras et al., 2019). This provides a principled explanation for why direct post hoc explanation of recurrent architectures remains technically delicate.

A more recent response to this difficulty is CatNet, which combines LSTM, SHAP derivatives, and Gaussian Mirrors to control the false discovery rate (FDR) in time-series feature selection. CatNet constructs mirror variables $f_t = \sigma(W_f x_t + U_f h_{t-1} + b_f)$1 and $f_t = \sigma(W_f x_t + U_f h_{t-1} + b_f)$2, trains an LSTM on augmented inputs, computes SHAP-based importance vectors, and forms a vector mirror statistic for thresholded FDR control. The method is explicitly designed to avoid instability from nonlinear or temporal correlations among features and is presented as naturally extensible to other sequential models such as RNN, GRU, and Attention-based models / Transformers (Han et al., 2024).

A common misconception is therefore that “LSTM-GA-SHAP” always means direct SHAP explanation of an LSTM tuned by GA. The literature does not support that assumption. Some pipelines explain tree models and transfer the resulting feature design to LSTM; some use surrogate models such as RandomForestRegressor or XGBoost to explain lagged features; and some omit GA entirely (Bhupatiraju et al., 29 Jul 2025, Rahimi et al., 8 Jul 2025).

6. Limitations, extensions, and research directions

Several limitations recur across the studies. First, the approach is sensitive to how recurrence is explained. The ERCOT study explicitly avoided SHAP on the final LSTM because of unstable recurrent explainers (Bhupatiraju et al., 29 Jul 2025), while the LRP study shows in detail why gated recurrence requires architecture-aware explanation rules (Arras et al., 2019). Second, GA adds computational overhead, especially when each candidate requires repeated model training, as in the campus electricity paper’s 5-fold cross-validation and multiple runs with different random seeds (Alba et al., 25 May 2026). Third, recursive or long-horizon forecasting compounds error accumulation, a point emphasized in both long-horizon energy studies (Alba et al., 25 May 2026, Bhupatiraju et al., 29 Jul 2025).

At the same time, the literature points to several clear directions for extension. The healthcare paper proposes real-time implementation, hybrid modeling, and reinforcement learning or PSO as future optimization alternatives (Rahimi et al., 8 Jul 2025). The campus electricity paper proposes shorter horizons, other universities and regions, institution-specific exogenous variables, and alternative ways to combine weaker predictors in WSB (Alba et al., 25 May 2026). CatNet proposes a path toward statistically controlled deep-sequence feature selection beyond LSTM, and its combination of SHAP-based importance with mirror-based testing suggests that interpretability and statistical validity need not remain separate design goals (Han et al., 2024).

More broadly, the literature indicates that the most stable and practically useful versions of the approach are those in which the roles of the three components are sharply distinguished: LSTM for temporal dependence, GA for search over hyperparameters or operational decisions, and SHAP for attribution, selection, or feature redesign. Where one of these roles is missing, the label may still be used, but only loosely. This suggests that the phrase “LSTM-GA-SHAP approach” is best understood not as a single algorithm, but as a configurable research pattern for combining recurrent forecasting, evolutionary optimization, and additive explanation in sequential decision problems (Bhupatiraju et al., 29 Jul 2025, Rahimi et al., 8 Jul 2025, Alba et al., 25 May 2026).