Table 1. Gap-filling methods hierarchy and parameters.

#	Category	Direction1	Variables2	Method	Architecture4	Key Parameters	Context Length3
1	Simple	–	Uni	Mean Imputation	Statistical	–	–
2	Simple	–	Uni	Median Imputation	Statistical	–	–
3	Local	–	Uni	Local Mean	Statistical	Window size = Variable	15 + gap + 15
4	Local	–	Uni	Local Median	Statistical	Window size = Variable	15 + gap + 15
5	Window	–	Uni	Linear Interpolation	Mathematical	–	10
6	Window	–	Uni	Polynomial Interpolation	Mathematical	Degree = 3	Variable
7	Window	–	Uni	B-spline Interpolation	Mathematical	–	Variable
8	Window	–	Uni	ARIMA Imputation	Statistical	Order=(1,0,0)	100
9	Autoreg	Uni	Uni	UniAR LSTM	Neural Network	Units = 64, Activation = tanh	32
10	Autoreg	Uni	Uni	UniAR GRU	Neural Network	Units = 64, Activation = tanh	32
11	Autoreg	Uni	Uni	UniAR RNN	Neural Network	Units = 64, Activation = tanh	32
12	Autoreg	Uni	Uni	UniAR CNN	Neural Network	Filters = 32, Kernel = 3	32
13	Autoreg	Uni	Uni	UniAR TCN	Neural Network	Filters = 32, Dilations=[1,2,4,8]	32
14	Autoreg	Uni	Uni	UniAR RF	Tree-based	Estimators = 50	32
15	Autoreg	Uni	Uni	UniAR XGB	Tree-based	Estimators = 50	32
16	Autoreg	Uni	Multi	UniAR XGB Multi	Tree-based	Estimators = 50, Features = 5	32
17	Seq2Seq	Uni	Uni	UniSeq2Seq LSTM	Neural Network	Units = 64, Bidirectional	32/32
18	Seq2Seq	Uni	Uni	UniSeq2Seq GRU	Neural Network	Units = 64, Bidirectional	32/32
19	Seq2Seq	Uni	Uni	UniSeq2Seq RNN	Neural Network	Units = 64, Bidirectional	32/32
20	Seq2Seq	Uni	Uni	UniSeq2Seq CNN	Neural Network	Filters = 32/64, Kernel = 3	32/32
21	Seq2Seq	Uni	Uni	UniSeq2Seq TCN	Neural Network	Filters = 64, Dilations=[1,2,4,8]	32/32
22	Seq2Seq	Uni	Uni	UniSeq2Seq RF	Tree-based	Estimators = 50	32/32
23	Seq2Seq	Uni	Uni	UniSeq2Seq XGB	Tree-based	Estimators = 50	32/32
24	Seq2Seq	Uni	Multi	UniSeq2Seq LSTM Multi	Neural Network	Units = 64, Bidirectional, Features = 5	32/32
25	Seq2Seq	Uni	Multi	UniSeq2Seq CNN Multi	Neural Network	Filters = 32, Kernel = 3, Features = 5	32/32
26	Seq2Seq	Uni	Multi	UniSeq2Seq RF Multi	Tree-based	Estimators = 50, Features = 5	32/32
27	Seq2Seq	Uni	Multi	UniSeq2Seq XGB Multi	Tree-based	Estimators = 50, Features = 5	32/32
28	Autoreg	Bi	Uni	LSTM Autoreg	Neural Network	Units = 64, Activation = tanh	32
29	Autoreg	Bi	Uni	GRU Autoreg	Neural Network	Units = 64, Activation = tanh	32
30	Autoreg	Bi	Uni	RNN Autoreg	Neural Network	Units = 64, Activation = tanh	32
31	Autoreg	Bi	Uni	CNN Autoreg	Neural Network	Filters = 32, Kernel = 3	32
32	Autoreg	Bi	Uni	TCN Autoreg	Neural Network	Filters = 32, Dilations=[1,2,4,8]	32
33	Autoreg	Bi	Uni	RF Autoreg	Tree-based	Estimators = 50	32
34	Autoreg	Bi	Uni	XGB Autoreg	Tree-based	Estimators = 50	32
35	Autoreg	Bi	Multi	XGB Autoreg Multi	Tree-based	Estimators = 50, Features = 5	32
36	Seq2Seq	Bi	Uni	LSTM Seq2Seq	Neural Network	Units = 64	32
37	Seq2Seq	Bi	Uni	GRU Seq2Seq	Neural Network	Units = 64	32
38	Seq2Seq	Bi	Uni	RNN Seq2Seq	Neural Network	Units = 64	32
39	Seq2Seq	Bi	Uni	CNN Seq2Seq	Neural Network	Filters = 32, Kernel = 3	32
40	Seq2Seq	Bi	Uni	TCN Seq2Seq	Neural Network	Filters = 32, Dilations=[1,2,4]	32
41	Seq2Seq	Bi	Uni	RF Seq2Seq	Tree-based	Estimators = 50	32
42	Seq2Seq	Bi	Uni	XGB Seq2Seq	Tree-based	Estimators = 50	32
43	Seq2Seq	Bi	Multi	XGB Seq2Seq Multi	Tree-based	Estimators = 50, Features = 5	32
44	Seq2Seq	Bi	Multi	RF Seq2Seq Multi	Tree-based	Estimators = 50, Features = 5	32
45	Dynamic	Uni	Uni	Dynamic Uni Seq2Seq XGB	Tree-based	Estimators = 50, Variable Context	Variable
46	Dynamic	Uni	Multi	Dynamic Multi Seq2Seq XGB	Tree-based	Estimators = 50, Features = 5, Variable Context	Variable

¹Direction: Uni = Unidirectional, Bi = Bidirectional

²Variables: Uni = Univariate (PM2.5 only), Multi = Multivariate (PM2.5 + meteorological and temporal features)

³Context Length: For bidirectional models, values represent pre/post gap context lengths

⁴Neural network models used Adam optimizer with learning rate = 0.001, trained for up to 30 epochs with early stopping (patience = 15). For all models, batch size = 32 was used during training