Automated feature engineering, AutoML, and decision-focused learning for improved energy consumption forecasting

Nasser Alkhulaifi

Figure 1.1 : Different potential applications of ECF. These applications rely on accurate ECF to optimise resource allocation, reduce costs, enhance grid stability, and support the transition to sustainable energy systems.

Figure 1.2 : Chapter structure and alignment with the research aim and objectives of this thesis. The abbreviations used in the figure are as follows: RO (Research Objective), ML (Machine Learning), AutoML (Automated Machine Learning), ECF (Energy Consumption Forecasting), FE (Feature Engineering), AFE (Automated Feature Engineering), FS (Feature Selection), and DFL (Decision-Focused Learning).

Figure 2.1 : Schematic representation of supervised ML adapted from [ 95 ] with modifications for clarity. It illustrates the relationship between training data

D=\{(X_{i},y_{i})\}_{i=1}^{N}

and the model learning process, where

X_{i}

represents the features,

y_{i}

represents the target, and

i

denotes the

i

-th data point in the dataset. The ideal target function

g(X_{i})

, which is unknown, maps features (i.e., inputs) to the target (i.e., output). During training, a parametric model produces predicted target

\hat{y}_{i}=h(X_{i},\phi)

, where

h

is the model function and

\phi

represents the learnable parameters.

Feature Extraction	Feature Type	Algorithm(s)	Energy System Application Context
	Calendar features: e.g., hour, day-of-week, month, holiday [ 203 , 195 , 175 , 88 ] .	ANN; KNN, RF, XGB, SVM, stacking [ 195 ] ; MTL-SVM [ 175 ] ; KNN ensemble [ 88 ] .	Campus buildings (heating) [ 195 ] ; Industrial parks [ 175 ] ; Public/commercial facilities [ 88 ] .
	Sin/Cos transformations: cyclical encoding (e.g., hour of day, day of week) using sine/cosine [ 101 , 131 ] .	Recurrent Inception CNN [ 101 ] ; Stacking with PCR [ 131 ] .	Industrial complexes [ 101 ] ; Office building (HQ) [ 131 ] .
	Lag-based (cyclic history): previous hour, day, week [ 195 , 34 , 63 , 113 ] .	KNN, RF, XGB; ANN [ 34 ] ; LSTM + attention [ 63 ] ; Deep MTL [ 113 ] .	Campus heating [ 195 ] ; Institutional cooling [ 34 ] ; Office electricity load [ 63 ] ; Commercial building [ 113 ] .
	Rolling-window statistics: e.g., moving min, max, mean, std [ 58 ] .	Elastic Net, GB Trees, RF, SVR, DNN, Auto-encoder [ 58 ] .	Educational building (chilled-water system) [ 58 ] .
	System/Site-specific: E.g., occupant count, indoor air temperature, humidity, zone loads; building metadata; lighting, equipment electricity [ 127 , 45 , 25 , 144 ] .	Bagging + genetic selection [ 127 ] ; SVR, MLP with Wavelet/Correlation [ 45 ] ; CNN-GRU + SHAP [ 25 ] ; 14 different ML methods [ 144 ] .	Occupancy-aware office [ 127 ] ; Office heating [ 45 ] ; Educational building (electricity) [ 25 ] ; UK residential (energy rating) [ 144 ] .
Feature Selection	Selection Method

Table 2.1 : Summary of different feature engineering approaches, inputs, and algorithms studied in the related work on energy consumption forecasting, along with the energy system/application contexts. See Table 2.2 for additional details on feature selection methods

Method	Specific technique	Description
Filter	Filter-driven with SVR kernels	Best features are chosen based on acquisition feasibility and performance scores using filter methods. These features are tested on datasets using support vector regression with radial basis and polynomial kernels [ 217 ] .
Embedded	Recursive feature elimination	This backward selection technique involves training a model with all variables, ranking them for importance, and iteratively removing the least important ones until no further reduction is possible [ 57 ] .
Wrapper	Genetic search	This method explores feature subsets using a genetic search strategy, focusing on predictive ability and minimising redundancy among features [ 127 ] .
Embedded	Data permutation-based importance	Optimal features are identified by assessing their role in enhancing predictive performance, with a focus on the impact of introducing irrelevant or noisy information [ 197 ] .
Hybrid	Two-stage filter-wrapper approach	Starts with a filter method to remove irrelevant features, reducing dimensionality without compromising accuracy. Then, a wrapper method conducts an exhaustive search for the most effective feature set, balancing accuracy and simplicity [ 214 ] .

Table 2.2 : Summary of different feature selection methods used in related work.

Figure 3.1 : Overview of literature on predicting energy consumption, including input features, FE, prediction methods, applications, areas of application, and identified research gaps for this work.

Figure 3.2 : Proposed ML pipeline for predicting electricity consumption, indoor temperature, and indoor humidity one week into the future in FDCS settings using weather data.

Figure 3.3 : 3D schematic representation of the investigated FDCSs. Key components indicated include (i) evaporator unit, responsible for absorbing heat from the storage space and maintaining low temperatures; (ii) condenser unit, essential for releasing the absorbed heat outside the storage room and condensing the refrigerant back into a liquid; (iii) an integrated meter collecting indoor temperature and humidity values; and (iv) a control panel that facilitates monitoring and adjusting the storage condition.

Figure 3.4 : Collected data from the two FDCS rooms plotted as a time series. Raw data points are represented by faint lines, while mean values are shown as solid lines. The mean is calculated by resampling the data into 24-hour blocks and taking the average value within each block, which smooths out short-term fluctuations and helps identify trends and anomalies.

Figure 3.5 : Hourly electricity, temperature, and humidity patterns for the two FDCSs. Each hourly reading is shown with reduced opacity, while mean hourly values are depicted in darker lines for clarity. The Coefficient of Variation (CV) quantifies relative variability. FDCS 1 shows higher electricity, temperature, and humidity variability (higher CV) than FDCS 2, indicating greater sensitivity to internal and/or external factors in FDCS 1 versus a more regulated environment in FDCS 2.

Variable	Statistics
	Mean	Std Dev	Min	25th Perc	50th Perc	75th Perc	Max
FDCS 1
Electricity Consumption (kWh)	922.22	677.51	106.01	178.35	958.15	1385.56	3613.5
Humidity Closer to Evaporator (%)	76.21	5.22	61.17	72.67	76.25	79.75	92.42
Temperature Closer to Evaporator ( $\mathrm{\SIUnitSymbolCelsius}$ )	12.16	1.84	9.5	10.67	11.42	13.5	16.92
Outdoor Temperature ( $\mathrm{\SIUnitSymbolCelsius}$ )	4.62	3.44	-3.30	2.04	4.44	6.80	13.76

Table 3.1 : Summary statistics of the collected data for electricity consumption, indoor humidity, temperature, and weather in both FDCSs.

Features	Type	Description
Weather data	Continuous	Outdoor Temperature ( $\mathrm{\SIUnitSymbolCelsius}$ ), Outdoor Dew Point ( $\mathrm{\SIUnitSymbolCelsius}$ ), Outdoor Wet Bulb Temperature ( $\mathrm{\SIUnitSymbolCelsius}$ ), Specific Humidity ( $\mathrm{g}\mathrm{/}\mathrm{k}\mathrm{g}$ ), Outdoor Relative Humidity (%), Precipitation (mm/hour), Surface Pressure (kPa), Wind Speed ( $\mathrm{m}\mathrm{/}\mathrm{s}$ ), Wind Direction (Degrees).
Temporal features	Integer Value	Hour of the day (0-23), day of the week (0-6), day of the month (1-31), the month of the year (1-12), weekday vs weekend (0-1)
Operation hours	Binary	The ’Is_Open’ feature serves as a binary indicator: 1 indicates that the storage is operating within working hours, while 0 indicates otherwise.
Cyclical features	Continuous	Sine and cosine transformation of temporal features (hour of day and day of week)
Time-lag features	Continuous	Previous values of the given target (e.g., electricity consumption from preceding hours/days)
Rolling-window statistical features	Continuous	Summary statistics (maximum, minimum, mean, kurtosis, skewness, and standard deviation) of a given target computed over a fixed-size window

Table 3.2 : Summary of input features used in this work.

Figure 3.6 : Correlation heatmaps of target variables against their time lags over one week for FDCS 1 and 2. The colour scale represents the Pearson correlation coefficients

Method	Type	Description	Hyperparameters
Correlation (F-Test)	Filter	Select top-k features using univariate LR tests based on F-statistics.	k = 20
Mutual Information	Filter	Chooses top-k features based on mutual information with the target, capturing non-linear relationships.	k = 20
Lasso Regularisation	Embedded	Uses L1 regularisation in Lasso regression to eliminate less important features by driving their coefficients to zero.	alpha = 0.01
Tree Importance (Extra Trees)	Embedded	Employs an ensemble of decision trees to rank features by importance, selecting those with higher importance.	n_estimators = 300
ElasticNet Regularisation	Embedded	Combines L1 and L2 regularisation, eliminating features with coefficients that shrink to zero.	alpha = 0.01
Recursive Feature Elimination (RFE)	Wrapper	Uses RFE with LR to recursively eliminate the least important features.	k = 20

Table 3.3 : Overview of FS methods used in the work (k = number of features to select).

Model	Description	Ref.
KNN Regression (KNR)	A non-parametric algorithm that predicts the target based on the average of the K closest training examples in the input feature space.	[ 184 ]
RF Regression (RFR)	An ensemble method that fits multiple decision trees on randomly sampled subsets of the data and combines their predictions.	[ 23 ]
XGB Regression (XGBR)	An ensemble method that trains decision trees sequentially, each time fitting the residual errors of the previous tree.	[ 29 ]
MLP Regressor	A feedforward ANN with multiple layers of nodes between input and output. Uses backpropagation to train the network weights and biases.	[ 2 ]
LSTM	A class of RNNs that can learn long-term dependencies by using a memory cell and three gating mechanisms.	[ 83 ]
MTL	A learning method where multiple tasks are handled concurrently, using a shared representation. By leveraging the similarities and variations across tasks, what is learnt for one task can aid in the learning of other tasks.	[ 26 ]

Table 3.4 : ML algorithms used in this work.

	Model	Best hyperparameters
FDCS 1	KNNR	’leaf_size’: 10, ’n_neighbours’: 10, ’weights’: ’distance’
	RFR	’max_depth’: 10, ’min_samples_leaf’: 2, ’min_samples_split’: 10, ’n_estimators’: 500
	XGBR	’learning_rate’: 0.1, ’max_depth’: 10, ’min_child_weight’: 2, ’n_estimators’: 500, ’subsample’: 0.8
	MLP	’activation’: ’relu’, ’alpha’: 0.0001, ’batch_size’: 16, ’hidden_layer_sizes’: (50, 100), ’learning_rate_init’: 0.01, ’max_iter’: 500, ’solver’: ’adam’
	LSTM	LSTM_Layers: [20,10], Optimizer: Adam, Learning_Rate: 0.01, Epochs: 100, batch_size: 32
	MTL	Base_Layers: 10, 40; Activation: ReLU; Optimizer: Adam; Learning_Rate: 0.005; Loss_Function: MSE; Epochs: 100; Batch_Size: 32; Loss_Weights: 1

Table 3.5 : Grid search results for hyperparameter tuning of models predicting electricity consumption. Training used a 4-core Intel CPU with 32GB RAM, Python 3.11, Scikit-learn, and TensorFlow frameworks.

Figure 3.7 : Model performance evaluation using MAE of the test set of different FS methods in FDCS 1.

Figure 3.8 : Model performance evaluation using MAE of the test set of different FS methods in FDCS 2.

Figure 3.9 : Comparison of the three lowest model errors in predicting electricity, temperature, and humidity in FDCS 1 and 2. Plots show actual versus predicted values on the test set.

Figure 3.10 : Analysis of the top ten features in terms of their impact on model output, as explained by SHAP for electricity, temperature, and humidity variables in FDCS 1.

Figure 3.11 : Analysis of the top ten features in terms of their impact on model output, as explained by SHAP for electricity, temperature, and humidity variables in FDCS 2.

Figure 3.12 : Variations of MAE in the test set occur when predicting electricity, temperature, and humidity in both FDCSs, as a function of increasing the training dataset size for the XGBR and RFR models.

Focus Area	Recommendations
Enhancement of input features	Use feature extraction (e.g., the hour of the day, and cyclical features).
FS techniques	Hybrid FS methods generally enhance model performance. Avoid wrapper methods if computational resources are limited.
Algorithms selection	Ensemble-based learning methods (XGBR, RFR) were superior to traditional ML, NN-based, and deep learning models in FDCS predictions. XGBR is particularly recommended for its computational efficiency.
Dataset collection	Contrary to popular belief, larger datasets were not necessary for accurate prediction in this work. Smaller datasets from real-world FDCS facilities may yield reliable predictions when enhanced with robust FE.

Table 3.6 : Summary of key recommendations for ML applications in FDCS systems, offering guidelines and insights to enhance the performance and efficiency of forecasting models in real-world applications.

Figure 4.1 : The proposed Algorithm 1 for generating temporal, sine, and cosine transform-derived features is illustrated in subfigure (A), while the proposed Algorithm 2 for generating lag and nested window features is illustrated in subfigure (B).

Figure 4.2 : Ordinal versus Fourier-based cyclical encoding of time features. (a) Ordinal hours (0-23) over one week show a discontinuity between 23 and 0. (b) First-order Fourier encoding,

\sin(2\pi h/24)

and

\cos(2\pi h/24)

, yields continuous cyclical features. (c) Unit-circle mapping assigns each hour unique coordinates (bijective, no information loss). (d) Distances between consecutive hours are constant (0.26), confirming the encoding preserves temporal resolution without smoothing, see Subsubsection 4.2.2.1 for more details.

Symbol	Definition
$\mathcal{D}$	Dataset composed of $N$ instances
$N$	Total number of instances in the dataset
$M$	Number of feature engineering functions
$t_{i}$	Timestamp associated with the $i$ -th instance
$y_{i}$	Target variable associated with the $i$ -th instance
$\mathbf{y}_{1:i-1}$	Historical target values available up to time $t_{i}$

Table 4.1 : Notations and definitions for all mathematical symbols, variables, and notation used throughout Section 4.2 .

Dataset	Description	Type	N	Resolution	Total Duration (days)	Ref.
AEP	Power consumption by American Electric Power	Uni	121,269	1h	5,053	[ 137 ]
Appliances	Energy consumption data from home appliances	Multi (27)	19,735	10m	138	[ 24 ]
CAISO_Elec	Electricity load by California ISO	Uni	26,304	1h	1,096	[ 128 ]
COMED	Energy usage by Commonwealth Edison	Uni	57,735	1h	2,406	[ 137 ]
DEOK	Energy consumption by Duke Energy OH/KY	Uni	57,735	1h	2,406	[ 137 ]
EKPC	Energy usage by East Kentucky Power	Uni	45,330	1h	1,889	[ 137 ]

Table 4.2 : Overview of the eighteen energy datasets used in this work. Column Dataset lists the dataset name. Column Description summarises the energy system and domain. Column Type indicates whether the dataset is univariate (Uni) or multivariate (Multi

(k)

, where

k

is the number of additional features, e.g., weather variables). Column N gives the total number of samples. Column Resolution shows the sampling interval (m = minutes, h = hours). Column Total Duration (days) gives the total time span covered by the dataset in whole days. Column Ref. cites the source.

Figure 4.3 : The energy datasets used in this work are depicted in histograms of the target variable with colour-coded bars representing normalised bin positions. Red and green dashed lines indicate the mean and standard deviation, respectively. Annotated statistics include coefficient of variation (CV: relative variability), skewness (distribution asymmetry), and kurtosis (tailedness). The colour gradient in the histogram bars represents the distribution of values from low (left) to high (right).

Figure 4.4 : Experimental design. See Subsection 4.3.3 for a detailed explanation of the experimental setup.

Figure 4.5 : Experimental design with TabPFN. See ”Special Case TabPFN” in Subsection 4.3.3 for a detailed explanation of this particular case.

	FE Methods
Dataset	AutoEnergy		FT		TSMin		TSEff		No.Feat.
	nRMSE	Time	nRMSE	Time	nRMSE	Time	nRMSE	Time	nRMSE	Time
AEP	0.0096	1585.90	0.0113	1576.84	0.1659	636.50	0.1737	4280.70	0.1856	0.00
Appliances	0.0870	44.33	0.0824	382.34	0.1831	177.06	0.1529	1411.01	0.1212	0.00
CAISO_Elec	0.0163	88.30	0.0232	211.91	0.2495	211.06	0.1074	1468.03	0.2036	0.00
COMED	0.0150	372.41	0.0151	558.55	0.1535	352.04	0.1579	2506.55	0.1905	0.00

Table 4.3 : Results obtained on all the test sets (N=18). nRMSE values and processing times in seconds (lower is better) for the proposed AutoEnergy method compared to other FE methods. Bold values highlight the lowest nRMSE and fastest processing time for each dataset. nRMSE gives more weight to larger errors due to the squaring operation, making it more sensitive to outliers. It is normalised by the range (max–min) of the target variable, allowing for fair comparison of forecasting error across datasets with different scales. Appendix B.1 in the appendices provides additional results of RMSE, MAE, MAPE and R 2 . Note that the TScomp configuration is not reported, as it proved computationally infeasible in our experiments (runs crashed after several hours of running due to excessive memory use). We include this note to highlight the practical computational limits of exhaustive FE approaches (e.g., TSComp) in real-world energy forecasting and to underscore the importance of considering computational feasibility and practicality in future work.

Figure 4.6 : Error analysis across test datasets. Each subplot shows the mean absolute error (MAE), computed as

\text{MAE}=\frac{1}{n}\sum_{i=1}^{n}|y_{i}-\hat{y}_{i}|

, between true and predicted values. The black line represents the MAE (lower is better), while the shaded areas and error bars indicate error variability. MAE provides an average magnitude of prediction error, treating all deviations equally and offering insight into prediction consistency that complements the scale-sensitive nRMSE in Table 4.3 .

Dataset	TabPFN	AutoEnergy	FT	TSEff	TSMin
Appliances	0.1412	0.0710	0.0712	0.1153	0.1497
FDCS_1	0.1549	0.0718	0.0778	0.0984	0.1822
FDCS_2	0.1504	0.1128	0.1278	0.1514	0.1620
Steel	0.0080	0.0081	0.0082	0.0135	0.0076
TCity	0.2691	0.0184	0.0212	0.0701	0.2299
UNICON	0.2226	0.0391	0.0422	0.2026	0.1963

Table 4.4 : nRMSE for TabPFN without FE versus with FE methods across datasets that satisfy the TabPFN constraints (

N\leq 10{,}000

,

D\leq 500

). Lower values are better. Boldface indicates the best results. Appendix B.2 in the appendices provides additional results of RMSE, MAE, MAPE and R 2 .

Figure 4.7 : Pearson correlation between dataset size (i.e., number of samples) and processing time for each method across all eighteen energy datasets. Markers represent individual datasets.

Dataset	Maximum Length
	1/4		1/3		1/2		1
	nRMSE	Time	nRMSE	Time	nRMSE	Time	nRMSE	Time
AEP	0.2241	1498.01	0.2229	1634.39	0.2227	1816.42	0.3124	2110.72
Appliances	0.2492	40.28	0.2447	44.35	0.2966	50.82	0.2429	63.48
CAISO_Elec	0.2002	85.50	0.1776	92.99	0.1669	104.12	0.1849	125.12
COMED	0.1519	367.54	0.1513	388.01	0.1498	439.48	0.1826	517.67

Table 4.5 : Sensitivity analysis results (nRMSE and processing time) for the time series fraction constraint defining the candidate range for nested window statistical feature generation, where only features with statistically significant Kendall’s tau correlation within this range are retained to generate the final model input features. The examined lengths are: quarter (1/4), one-third (1/3), half (1/2), and full sequence length (1). This experiment evaluates the impact of varying maximum rolling window fractions on predictive model performance and efficiency, using solely the nested window features as inputs to isolate their standalone contributions and inform optimal parameter selection. See Subsubsection 4.2.2.2 for more details.

Figure 4.8 : Comparison of the top 20 most important features across datasets, grouped into three categories: Temporal, Sine and Cosine Transform-Derived Features (TSC; Algorithm 1 ), Lags and Nested Window Features (LNW; Algorithm 2 ), and Other (e.g., weather). The left panel shows feature counts, and the right panel shows cumulative normalised importance (proportion of total importance). LNW features contribute most, consistent with their ability to capture temporal dependencies in energy consumption, including lagged effects such as thermal inertia [ 126 ] and multi-scale behaviour via nested windows. More analysis is provided in Subsection 4.4.4 , while the full feature-importance scores are provided in the supplementary material (see Appendix C in the appendices).

Figure 4.9 : Comparisons between the proposed FE algorithm and other FE methods using the Friedman Average Ranking (lower is better).

nRMSE
Comparison	Test Statistic	Std. Error	Std. Test Statistic	Sig.	Adj. Sig. a
AutoEnergy-FT	-0.917	0.527	-1.739	0.082	0.820
AutoEnergy-TSEff	-2.556	0.527	-4.849	$<$ 0.001	0.000
AutoEnergy-TSMin	-3.028	0.527	-5.745	$<$ 0.001	0.000
AutoEnergy-No.Feat.	-3.222	0.527	-6.114	$<$ 0.001	0.000
Processing Time

Table 4.6 : Friedman test results for nRMSE and processing time comparisons.

nRMSE
Metric	Comparison	Z	Asymp. Sig. (2-tailed)
	FT - AutoEnergy	-3.202 b	0.001
	TSEff - AutoEnergy	-3.724 b	<0.001
	TSMin - AutoEnergy	-3.724 b	<0.001
	No.Feat. - AutoEnergy	-3.724 b	<0.001
Processing Time

Table 4.7 : Wilcoxon Signed Ranks test results comparing AutoEnergy with other FE methods for nRMSE and processing time.

Figure 5.1 : The proposed framework jointly forecasts next-day electricity prices and demand while optimising the downstream BESS scheduling task. It uses automated FE to enrich the dataset with domain-specific features via AutoEnergy [ 7 ] (details in Subsection 5.2.4 ). The goal is to reduce domain-knowledge requirements and improve decision quality under data scarcity. While the problem formulation (see Subsection 5.2.2 ) is BESS-specific, the overall methodology has the potential to be adapted to other similar energy storage systems by revising the constraints, variables, and objectives.

Set	Definition
$T$	Set of (time) intervals
$max\_charge$	Maximum energy that can be added to the battery in an interval
$max\_discharge$	Maximum energy that can be drained from the battery in an interval
$p_{t}$	Price of 1 kWh energy at interval $t$	$\forall t\in T$
$d_{t}$	Energy demand at interval $t$	$\forall t\in T$
$e_{0}$	Initial battery level (kWh)

Table 5.1 : Notations

Dec. Var.	Definition
$e_{t}$	Amount of energy (kWh) in the battery at the end of interval $t$
$g_{t}$	Amount of energy (kWh) provided from grid to the property at interval $t$
$b_{t}$	Amount of energy (kWh) provided from battery to the property at interval $t$
$c_{t}$	Amount of energy (kWh) charged to battery at interval $t$
$z_{t}$	1, if energy is being sent from battery to the property at interval $t$
	0, otherwise

Table 5.2 : Decision Variables

Figure 5.2 : The dataset used in this work is depicted in histograms with colour-coded bars representing normalised bin positions. Red and green dashed lines indicate the mean and standard deviation, respectively. Annotated statistics include coefficient of variation (CV: relative variability), skewness (distribution asymmetry), and kurtosis (tailedness). The colour gradient in the histogram bars represents the distribution of values from low (left) to high (right).

Figure 5.3 : Heatmaps of average hourly household electricity demand (top) and price (bottom) by day of week. Demand displays consistent peaks in the early morning (around breakfast), midday (lunch), and most prominently between 17:30 and 20:00 (dinner time), reflecting typical residential consumption patterns. Price peaks are concentrated between 16:00 and 19:30. Both demand and price are lowest during the early morning hours (01:00 to 05:00), reflecting reduced residential activity and system load during overnight periods.

Parameter	Value
Planning horizon $T$ [intervals]	$24$
Interval duration $\Delta t$ [hours]	$1$
Usable capacity $E_{\max}$ [kWh]	$50$
Minimum SoC fraction $\mathrm{SoC}_{\min}$	$0.10$
Max charge per interval $max\_charge$ [kWh]	$5.0$
Max discharge per interval $max\_discharge$ [kWh]	$4.5$

Table 5.3 : Summary of parameters and values used in the experiment. See Subsection 5.3.2 for BESS configuration details and Subsection 5.3.3 for the experimental procedure.

Figure 5.4 : Power flow diagram, where: A) power flows from the grid to charge the battery, which then supplies the property via inverters performing DC–AC conversion to meet energy demand; and B) power flows directly from the grid to the property when the battery capacity is insufficient.

Figure 5.5 : Experimental design. See Subsection 5.3.3 for a detailed explanation of the experimental procedure.

Hyperparameter	Search Space	Best Hyperparameter
		PTO		SPO +		DBB
		No AFE	AFE	No AFE	AFE	No AFE	AFE
Number of Layers	$\in\{1,\,2\}$	2	2	2	2	2	1
Epochs	$\in\{10,\,20,\,30\}$	30	30	30	30	30	30
Learning Rate	$\in\{10^{-3},10^{-5}\}$	$10^{-3}$	$10^{-3}$	$10^{-5}$	$10^{-5}$	$10^{-3}$	$10^{-3}$

Table 5.4 : Hyperparameter search space and the corresponding best hyperparameters for the ANN models, determined through grid search on the validation dataset. To avoid over-parameterisation, the search space is constrained to the most critical parameters identified in related work [ 147 ] . However, it is worth mentioning that this may have restricted the discovery of optimal hyperparameter combinations that could further enhance model performance .

Figure 5.6 : Test regret across methods with and without AFE. (A) Box plots of test regret across all experimental runs and test days for PTO and DFL methods (SPO + and DBB). The distributions reflect performance variability across multiple independent experimental runs and test days, providing a comprehensive view of method robustness and central tendency. (B) Daily mean test regret over the fourteen test days. (C) Mean test regret across test days for each method variant. (D) Standard deviation of test regret across test days. These results support DFL being superior in this experimental setting when combined with AFE (as measured by regret), but they do not imply that DFL is universally superior across all datasets, objectives, or optimisation problems.

Date	Test Set Regrets
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.0620	0.0780	0.0079	0.0611	0.0832	0.1247
12 Feb	0.1234	0.1669	0.0369	0.0845	0.1440	0.1858
13 Feb	0.1573	0.1968	0.0205	0.1222	0.1685	0.2243
14 Feb	0.1856	0.2363	0.0469	0.1736	0.2101	0.2678

Table 5.5 : Daily test regrets (lower is better) averaged across ten independent runs to improve the robustness and reliability of the results (see supplementary materials for detailed results of each of the ten runs across the test set days). The lowest mean and standard deviation values are bolded to indicate the best performance. The comparison includes PTO and DFL methods (SPO + , DBB), each evaluated with and without AFE (see Subsection 5.3.3 for more details on the experimental procedure). Appendix B provides detailed daily BESS optimisation schedules for the best run of each day using the best-performing method.

Figure 5.7 : Feature importance across all experimental runs and test days. Each subplot corresponds to one method and reports the mean SHAP importance per feature. Raw inputs (“No AFE”, e.g., weather) are compared with engineered features (“AFE”, e.g., lags, rolling-window statistics, and calendar indicators). Markers (colour and shape) denote feature categories; each point is a feature, with x-axis jitter to reduce overlap and show within-category dispersion. More detailed analysis is provided in Subsection 5.4.3 .

Figure 5.8 : Result of comparisons on the test set between the PTO and DFL (SPO + and DBB) methods, with and without AFE, using the Friedman average ranking (lower is better). The significance level is

0.05

.

Method 1 vs Method 2	Test Statistic	Std. Test Statistic	Sig.	Adj. Sig. ${}^{\text{a}}$
SPO + (AFE) vs SPO + (No AFE)	-1.143	-1.616	.106	1.000
SPO + (AFE) vs PTO (AFE)	2.000	2.828	.005	.070
SPO + (AFE) vs DBB (AFE)	-3.357	-4.748	$<.001$	.000
SPO + (AFE) vs PTO (No AFE)	3.500	4.950	$<.001$	.000
SPO + (AFE) vs DBB (No AFE)	-5.000	-7.071	$<.001$	.000
SPO + (No AFE) vs PTO (AFE)	0.857	1.212	.225	1.000

Table 5.6 : Results of pairwise comparisons on the test set between the PTO and DFL (SPO + and DBB) methods, with and without AFE, using the Wilcoxon signed-rank test. Each row tests the null hypothesis that the distributions of Sample 1 (i.e., Method 1) and Sample 2 (i.e., Method 2) are the same. Asymptotic p-values (2-sided tests) are displayed. The significance level is

0.05

. P-values below the significance level indicate rejection of the null hypothesis. Values reported as ¡ .001 signify p-values less than 0.001, providing strong statistical evidence. a Significance values have been adjusted by the Bonferroni correction for multiple tests.

Chapter	Title of Article	Novelty
Chapter 3	Machine Learning Pipeline for Energy and Environmental Prediction in Cold Storage Facilities	Proposed a comprehensive ML pipeline for ECF capable of one-week-ahead hourly forecasting and evaluated under data-constrained settings . The pipeline was validated on two novel UK-based cold storage facility datasets. Systematic evaluation of eight FS techniques with extensive analysis of domain knowledge impact, feature importance, and dataset size implications in ECF applications. This provides a baseline for later automation of FE.
Appendix A	Exploring Automated Feature Engineering for Energy Consumption Forecasting with AutoML	Developed an ECF-specific AFE prototype to reduce the domain expertise needed for FE. Reported performance gains in the evaluated settings across four AutoML frameworks (AutoGluon, H2O, TPOT, and FLAML).
Chapter 4	AutoEnergy: An Automated Feature Engineering Algorithm for Energy Consumption Forecasting with AutoML	Introduced AutoEnergy , a novel, fully automated FE algorithm tailored for ECF that generates interpretable features from timestamps and past consumption values through rule-based transformations, integrating with AutoML to reduce human intervention. Evaluated across eighteen diverse real-world energy datasets representing different energy systems, AutoEnergy showed promising generalisability across the evaluated datasets while typically reducing error and achieving competitive processing time compared to existing methods, with integration with the TabPFN foundation model on eligible datasets .
Chapter 5	Decision-Focused Learning Enhanced by Automated Feature Engineering for Energy Storage Optimisation	Leveraged AutoEnergy to support the downstream task and study the nascent DFL, validated through a novel BESS dataset in real-world settings. On the evaluated case study, results show that incorporating AutoEnergy further improves DFL decision quality by achieving 22.9-56.5% lower operating costs compared to the same models without it.

Table 6.1 : Summary of publications arising from this thesis.

Dataset	Description	Type	Granularity	Samples	Mean	Std	Min	Max	Ref.
Appliances	Home appliances energy consumption	Multivariate (27 features)	10mins	19,735	97.69	102.5	10.0	1,080.0	[ 24 ]
TCity	Power consumption of the city of Tetouan, Morocco	Multivariate (5 features)	10mins	52,416	32,344.9	7,130.5	13,895.7	52,204.4	[ 159 ]
WindT	SCADA system data for wind turbines	Multivariate (2 features)	10mins	50,530	1,307.6	1,312.4	2.4	3,618.7	[ 169 ]
AEP	American Electric Power consumption	Univariate	Hourly	121,273	15,499.6	2,591.3	9,581.0	25,695.0	[ 137 ]
PJME	PJM East Region power consumption	Univariate	Hourly	144,366	32,080.5	6,463.8	14,544.0	62,009.0	[ 137 ]
PJMW	PJM West Region power consumption	Univariate	Hourly	143,206	5,602.4	979.1	487.0	9,594.0	[ 137 ]

Table A.1 : Datasets used in this study representing different energy patterns in residential buildings, renewable energy, and regional energy consumption. The multivariate datasets mean they have additional features (e.g., weather).

Figure A.1 : Proposed FE algorithm and experimental setup.

Dataset	H2O		TPOT		AutoGluon		FLAML
	with	without	with	without	with	without	with	without
Appliances	0.0725	0.2398	0.0777	0.1054	0.0876	0.1109	0.0770	0.1507
TCity	0.0118	0.2274	0.0149	0.2064	0.0122	0.2331	0.0131	0.2418
WindT	0.0442	0.1383	0.0595	0.1419	0.0608	0.1445	0.0497	0.1395
AEP	0.0107	0.2637	0.0173	0.2120	0.0101	0.2421	0.0104	0.2649
PJME	0.0072	0.2832	0.0142	0.1769	0.0066	0.1797	0.0082	0.2894

Table A.2 : Comparison of nRMSEs (lower is better) of AutoML methods across different datasets with vs. without using the proposed FE method. The boldface is the lowest (i.e., best) nRMSE value.

	H2O	TPOT	AutoGluon	FLAML
Z value a	-2.934	-2.936	-2.936	-2.936
P-value	0.003	0.003	0.003	0.003

Table A.3 : Results of pairwise comparisons of each AutoML using the Wilcoxon Test.

Figure A.2 : Average Friedman Ranks of all AutoML methods. N=11, P

<

.001, Chi-Square = 64.636.

Dataset	AutoGluon
	All	Time-based & cyclic	Lags & rolling-stats
Appliances	0.0876	0.1090	0.0929
TCity	0.0122	0.1540	0.0124
WindT	0.0608	0.1250	0.0386
AEP	0.0101	0.1471	0.0122
PJME	0.0066	0.1190	0.0081

Table A.4 : Comparison of nRMSEs (lower is better) using various input features across different datasets.

	FE Methods
Dataset	AutoEnergy		FT		TSMin		TSEff		No.Feat.
	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE
AEP	125.88	90.310	148.90	112.09	2515.3	1962.9	2633.5	2054.9	2445.8	1959.3
Appliances	72.205	26.727	68.366	23.436	151.98	78.951	126.89	53.078	100.58	69.336
CAISO_Elec	573.73	368.86	818.03	592.44	8785.4	7376.6	3782.1	2943.9	7171.6	5677.0
COMED	64.451	28.250	64.819	31.482	658.33	534.51	676.95	541.09	816.74	664.80

Table B.1 : RMSE and MAE results for various FE methods using AutoGluon across all test sets. Lower is better. Bold indicates best results. These results are complementary to those presented in Table 4.3 .

	FE Methods
Dataset	AutoEnergy		FT		TSMin		TSEff		No.Feat.
	MAPE%	R2	MAPE%	R2	MAPE%	R2	MAPE%	R2	MAPE%	R2
AEP	0.62000%	0.99730	0.77000%	0.99630	12.940%	0.040100	13.440%	-0.052300	13.660%	-0.0014000
Appliances	18.030%	0.37090	14.350%	0.43610	82.030%	-1.7868	46.070%	-0.94280	90.740%	-0.22070
CAISO_Elec	1.9900%	0.99010	3.0200%	0.97990	32.960%	-1.3214	14.770%	0.56980	31.570%	-0.54690
COMED	1.0500%	0.98920	1.1500%	0.98910	18.270%	-0.15890	17.920%	-0.22540	21.640%	-0.72930

Table B.2 : MAPE(%) and

R^{2}

Results for various FE methods using AutoGluon across all test sets. Lower MAPE and higher

R^{2}

are better. Bold indicates best results. These results are complementary to those presented in Table 4.3 .

	Methods
Dataset	TabPFN		TabPFN_AutoEnergy		TabPFN_FT		TabPFN_TSEff		TabPFN_TSMin
	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE
Appliances	117.18	102.10	58.891	24.478	59.070	24.179	95.739	74.172	124.23	111.26
FDCS_1	444.96	321.52	206.34	127.91	223.44	153.03	282.83	200.28	523.57	414.20
FDCS_2	443.46	343.62	332.63	238.43	376.95	277.53	446.51	348.85	477.77	373.33
Steel	4.2997	1.8133	4.3478	1.9157	4.3778	2.1459	7.2560	4.1050	4.0600	1.7699

Table B.3 : RMSE and MAE results for various FE methods using TabPFN across all test sets. Bold indicates the best per row and metric. These results are complementary to those presented in Table 4.4 .

	Methods
Dataset	TabPFN		TabPFN_AutoEnergy		TabPFN_FT		TabPFN_TSEff		TabPFN_TSMin
	MAPE%	R2	MAPE%	R2	MAPE%	R2	MAPE%	R2	MAPE%	R2
Appliances	146.00%	-0.57350	19.610%	0.60250	18.830%	0.60010	97.080%	-0.050400	161.76%	-0.76870
FDCS_1	88.250%	0.50360	17.600%	0.89330	30.150%	0.87480	43.360%	0.79940	179.56%	0.31270
FDCS_2	46.260%	0.34480	28.400%	0.63130	37.090%	0.52660	46.650%	0.33570	52.540%	0.23940
Steel	1.7000%	0.99880	2.1500%	0.99880	2.6400%	0.99880	5.4200%	0.99670	1.7900%	0.99900

Table B.4 : MAPE(%) and

R^{2}

results for various FE methods using TabPFN across all test sets. Bold indicates the best per row and metric. These results are complementary to those presented in Table 4.4 .

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	0.023	0.003	0.000001	7	0.028	0.019
y_window_2_min	0.012	0.003	0.000021	7	0.016	0.008
y_lag_2	0.008	0.002	0.000057	7	0.011	0.004
y_window_2_max	0.007	0.003	0.000324	7	0.012	0.003
y_window_2_mean	0.007	0.003	0.000396	7	0.011	0.003
y_window_3_min	0.005	0.003	0.002797	7	0.009	0.001

Table C.1 : Feature Importance Analysis for Appliances Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	6472.919	61.023	0.000001	10	6535.631	6410.207
y_window_2_mean	274.030	4.320	0.000001	10	278.470	269.590
y_window_2_min	133.148	3.328	0.000001	10	136.568	129.728
hour	84.677	2.237	0.000001	10	86.977	82.378
y_window_2_max	76.294	2.970	0.000001	10	79.347	73.242
hour_sin	66.290	3.312	0.000001	10	69.693	62.886

Table C.2 : Feature Importance Analysis for TCity Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
Wind Speed	1089.044	14.551	0.000001	4	1131.540	1046.547
y_lag_1	221.907	4.591	0.000001	4	235.314	208.500
y_window_2_mean	39.566	0.603	0.000001	4	41.327	37.804
y_window_2_min	28.250	0.786	0.000003	4	30.546	25.955
y_window_2_max	11.408	0.761	0.000041	4	13.632	9.185
month_sin	8.698	0.838	0.000122	4	11.146	6.250

Table C.3 : Feature Importance Analysis for WindT Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	2959.685	61.399	0.000001	4	3138.997	2780.374
hour_cos	164.392	7.553	0.000013	4	186.449	142.335
y_window_2_mean	91.869	7.072	0.000063	4	112.522	71.216
hour	86.483	2.501	0.000003	4	93.786	79.180
y_window_2_min	84.227	6.214	0.000055	4	102.376	66.078
hour_sin	80.137	5.265	0.000039	4	95.513	64.761

Table C.4 : Feature Importance Analysis for AEP Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	9328.038	160.553	0.000001	5	9658.618	8997.458
y_lag_2	1072.594	35.865	0.000001	5	1146.442	998.747
hour_cos	627.283	34.718	0.000001	5	698.767	555.798
y_window_2_std	260.277	19.365	0.000004	5	300.150	220.404
hour	206.776	13.390	0.000002	5	234.347	179.206
hour_sin	187.717	14.653	0.000004	5	217.887	157.546

Table C.5 : Feature Importance Analysis for PJME Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	1178.745	3.675	0.000001	4	1189.479	1168.011
hour_cos	87.534	2.280	0.000002	4	94.192	80.876
y_window_2_mean	27.700	0.239	0.000001	4	28.397	27.002
hour_sin	25.967	0.836	0.000005	4	28.408	23.527
hour	25.166	0.589	0.000002	4	26.886	23.445
y_window_2_max	19.485	0.249	0.000001	4	20.211	18.759

Table C.6 : Feature Importance Analysis for PJMW Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	666.802	7.442	0.000001	7	677.230	656.374
hour_cos	39.205	2.043	0.000001	7	42.067	36.342
y_window_2_mean	26.871	0.558	0.000001	7	27.654	26.089
y_window_2_min	17.563	0.665	0.000001	7	18.495	16.631
y_lag_2	10.928	0.733	0.000001	7	11.955	9.901
hour	10.744	0.609	0.000001	7	11.597	9.890

Table C.7 : Feature Importance Analysis for COMED Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	1696.252	17.912	0.000001	4	1748.564	1643.940
hour_cos	115.522	7.444	0.000037	4	137.263	93.781
y_lag_2	66.634	4.003	0.000030	4	78.326	54.942
hour	55.829	9.527	0.000667	4	83.652	28.005
hour_sin	38.892	3.179	0.000075	4	48.176	29.608
month_cos	20.839	1.478	0.000049	4	25.157	16.522

Table C.8 : Feature Importance Analysis for FE Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	2666.381	53.292	0.000001	4	2822.018	2510.743
hour	259.923	7.318	0.000003	4	281.295	238.550
hour_cos	155.461	7.185	0.000014	4	176.445	134.477
y_window_2_max	112.643	10.584	0.000113	4	143.551	81.734
y_window_2_mean	85.701	8.234	0.000121	4	109.749	61.652
y_window_2_std	84.605	5.606	0.000040	4	100.978	68.232

Table C.9 : Feature Importance Analysis for NI Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	666.802	7.442	0.000001	7	677.230	656.374
hour_cos	39.205	2.043	0.000001	7	42.067	36.342
y_window_2_mean	26.871	0.558	0.000001	7	27.654	26.089
y_window_2_min	17.563	0.665	0.000001	7	18.495	16.631
y_lag_2	10.928	0.733	0.000001	7	11.955	9.901
hour	10.744	0.609	0.000001	7	11.597	9.890

Table C.10 : Feature Importance Analysis for DEOK Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	437.837	3.513	0.000001	10	441.447	434.227
hour_cos	18.492	0.615	0.000001	10	19.124	17.859
y_window_2_min	17.231	0.517	0.000001	10	17.762	16.700
hour_sin	14.479	0.296	0.000001	10	14.783	14.175
y_window_2_mean	14.187	0.393	0.000001	10	14.590	13.783
y_window_2_max	10.753	0.341	0.000001	10	11.103	10.403

Table C.11 : Feature Importance Analysis for EKPC Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
Lagging_Current_Reactive.Power_kVarh	138.335	1.550	0.000001	10	139.928	136.742
Lagging_Current_Power_Factor	71.301	0.840	0.000001	10	72.164	70.438
Leading_Current_Power_Factor	26.245	0.690	0.000001	10	26.954	25.536
y_lag_1	7.810	0.187	0.000001	10	8.003	7.618
Leading_Current_Reactive_Power_kVarh	2.586	0.196	0.000001	10	2.787	2.385
y_window_2_max	0.748	0.036	0.000001	10	0.784	0.711

Table C.12 : Feature Importance Analysis for Steel Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	2.608	0.046	0.000001	10	2.655	2.560
hour_cos	0.365	0.015	0.000001	10	0.380	0.349
y_window_2_mean	0.231	0.009	0.000001	10	0.241	0.221
hour_sin	0.155	0.010	0.000001	10	0.165	0.144
y_window_2_min	0.152	0.007	0.000001	10	0.160	0.145
hour	0.115	0.012	0.000001	10	0.128	0.103

Table C.13 : Feature Importance Analysis for UNICON Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	805.529	6.196	0.000010	3	841.032	770.026
y_window_2_max	80.172	1.050	0.000029	3	86.187	74.156
y_window_2_mean	66.534	1.053	0.000042	3	72.567	60.501
y_window_2_min	55.862	0.858	0.000039	3	60.781	50.943
hour	22.948	0.249	0.000020	3	24.378	21.519
hour_sin	15.737	0.878	0.000518	3	20.766	10.708

Table C.14 : Feature Importance Analysis for Victoria Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	6180.657	53.102	0.000001	10	6235.229	6126.085
y_lag_2	973.309	11.782	0.000001	10	985.418	961.201
hour_sin	673.672	9.721	0.000001	10	683.662	663.682
y_lag_23	504.967	8.934	0.000001	10	514.148	495.786
hour_cos	404.239	10.045	0.000001	10	414.562	393.915
y_window_2_std	114.964	3.910	0.000001	10	118.982	110.946

Table C.15 : Feature Importance Analysis for CAISO_Elec Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_1	0.068	0.003	0.000001	10	0.071	0.065
hour_sin	0.008	0.001	0.000001	10	0.009	0.007
hour	0.006	0.001	0.000001	10	0.007	0.006
hour_cos	0.004	0.000	0.000001	10	0.004	0.003
y_lag_2	0.002	0.000	0.000001	10	0.003	0.002
y_window_11_mean	0.002	0.000	0.000001	10	0.002	0.002

Table C.16 : Feature Importance Analysis for Solar_Home Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_24	176.012	13.971	0.000001	10	190.369	161.655
hour_cos	52.557	4.317	0.000001	10	56.994	48.120
y_lag_48	16.092	2.042	0.000001	10	18.190	13.994
hour_sin	11.471	1.721	0.000001	10	13.239	9.702
hour	9.078	1.740	0.000001	10	10.867	7.290
y_lag_72	8.721	1.203	0.000001	10	9.958	7.484

Table C.17 : Feature Importance Analysis for FDCS_1 Dataset

Feature	Importance	Std Dev	p-value	n	p99 High	p99 Low
y_lag_3	111.814	7.866	0.000001	10	119.898	103.730
y_lag_18	47.902	4.638	0.000001	10	52.668	43.136
y_lag_19	26.304	4.133	0.000001	10	30.551	22.057
hour_cos	21.691	5.233	0.000001	10	27.068	16.313
hour_sin	16.210	4.318	0.000001	10	20.647	11.772
y_lag_20	14.978	1.293	0.000001	10	16.308	13.649

Table C.18 : Feature Importance Analysis for FDCS_2 Dataset

Date	Test Set Regrets — Run 1
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.0460	0.1005	0.0022	0.0292	0.0252	0.1617
12 Feb	0.1518	0.1661	0.0252	0.0643	0.0901	0.2267
13 Feb	0.2054	0.2000	0.0050	0.0918	0.1022	0.2584
14 Feb	0.2161	0.2286	0.0378	0.1566	0.0963	0.3694

Table D.1 : Daily test regrets for Run 1.

Date	Test Set Regrets — Run 2
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.0351	0.0647	0.0148	0.0437	0.0455	0.1620
12 Feb	0.0886	0.1312	0.0544	0.0808	0.0551	0.2263
13 Feb	0.1096	0.1917	0.0118	0.0727	0.0576	0.2750
14 Feb	0.1404	0.2171	0.1067	0.1041	0.1247	0.3127

Table D.2 : Daily test regrets for Run 2.

Date	Test Set Regrets — Run 3
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.0541	0.0877	0.0044	0.0453	0.0755	0.1527
12 Feb	0.0876	0.1597	0.0329	0.0371	0.1857	0.2540
13 Feb	0.1369	0.1321	0.0166	0.0954	0.1669	0.2749
14 Feb	0.1310	0.2120	0.0346	0.1521	0.2073	0.3706

Table D.3 : Daily test regrets for Run 3.

Date	Test Set Regrets — Run 4
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.1763	0.0823	0.0065	0.0494	0.1317	0.0730
12 Feb	0.2184	0.1498	0.0334	0.0785	0.1565	0.2161
13 Feb	0.2814	0.1939	0.0286	0.0912	0.2282	0.1393
14 Feb	0.3800	0.2134	0.0848	0.1494	0.2801	0.2521

Table D.4 : Daily test regrets for Run 4.

Date	Test Set Regrets — Run 5
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.0426	0.0965	0.0117	0.1344	0.1952	0.0534
12 Feb	0.0691	0.2341	0.0186	0.0891	0.2858	0.0611
13 Feb	0.0901	0.2732	0.0477	0.0827	0.3764	0.0994
14 Feb	0.1358	0.3106	0.0394	0.2744	0.3711	0.1486

Table D.5 : Daily test regrets for Run 5.

Date	Test Set Regrets — Run 6
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.0593	0.0884	0.0041	0.0138	0.0888	0.1592
12 Feb	0.0949	0.2681	0.0455	0.0927	0.1472	0.1502
13 Feb	0.1030	0.2631	0.0308	0.0800	0.2264	0.2567
14 Feb	0.1240	0.2599	0.0680	0.1366	0.1880	0.2247

Table D.6 : Daily test regrets for Run 6.

Date	Test Set Regrets — Run 7
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.0716	0.0491	0.0051	0.1292	0.0520	0.1019
12 Feb	0.2049	0.1306	0.0172	0.1614	0.1301	0.1913
13 Feb	0.2482	0.2323	0.0026	0.1848	0.1023	0.2722
14 Feb	0.2851	0.2162	0.0121	0.2314	0.2267	0.2332

Table D.7 : Daily test regrets for Run 7.

Date	Test Set Regrets — Run 8
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.0305	0.0459	0.0050	0.0405	0.0898	0.0975
12 Feb	0.0408	0.1060	0.0168	0.0785	0.1520	0.1342
13 Feb	0.0617	0.1135	0.0080	0.1805	0.1569	0.1610
14 Feb	0.0864	0.2115	0.0072	0.1535	0.2392	0.1871

Table D.8 : Daily test regrets for Run 8.

Date	Test Set Regrets — Run 9
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.0501	0.0789	0.0205	0.0521	0.0518	0.1518
12 Feb	0.1817	0.1710	0.0914	0.0695	0.1261	0.2286
13 Feb	0.1752	0.1990	0.0511	0.1262	0.1794	0.2573
14 Feb	0.2021	0.2713	0.0723	0.1508	0.1832	0.2796

Table D.9 : Daily test regrets for Run 9.

Date	Test Set Regrets — Run 10
	PTO		SPO +		DBB
	AFE	No AFE	AFE	No AFE	AFE	No AFE
11 Feb	0.0547	0.0858	0.0051	0.0734	0.0763	0.1337
12 Feb	0.0959	0.1523	0.0339	0.0928	0.1110	0.1692
13 Feb	0.1616	0.1696	0.0026	0.2173	0.0892	0.2493
14 Feb	0.1555	0.2224	0.0066	0.2267	0.1844	0.3002

Table D.10 : Daily test regrets for Run 10.

Figure E.1 : Daily BESS optimisation results showing best-performing methods per day. Green circles indicate battery charging, red circles show discharging, and orange background bars represent direct grid supply. The dashed blue line shows predicted electricity prices (right y-axis).

Figure E.2 : Daily BESS optimisation results showing best-performing methods per day. Green circles indicate battery charging, red circles show discharging, and orange background bars represent direct grid supply. The dashed blue line shows predicted electricity prices (right y-axis).

Automated feature engineering, AutoML, and decision-focused learning for improved energy consumption forecasting

Abstract

Figures & tables

Explore similar work

FETS Benchmark: Foundation Models Outperform Dataset-specific Machine Learning in Energy Time Series Forecasting

Electricity Consumption Forecasting: An Approach Using Cooperative Ensemble Learning with SHapley Additive exPlanations

Distributed JEPA: A Self-Supervised Framework for Energy Forecasting